Method for predicting multi-layer commingling production capacity of tight sandstone gas reservoir

Abstract

The invention provides a compact sandstone gas reservoir multi-layer commingling production capacity prediction method, which comprises the following steps of: constructing a graph data model containing operation variables, static parameters and dynamic data based on a Spearman correlation coefficient, updating node characteristics in real time through an Internet of Things sensor, and correcting an adjacent matrix weight by adopting a dynamic graph mechanism; s2, extracting multi-modal data features by using a CNN network model and an LSTM network model, inputting the multi-modal data features into an OV-GCN model after PCA dimension reduction, realizing graph structure modeling in combination with GAT and Transform, and outputting predicted values of productivity distribution, moisture content and interlayer interference; and S3, inputting a prediction result into an NSGA-II multi-target optimization model, and generating a fracturing parameter Pareto optimal solution set by taking productivity maximization, cost minimization and interference balance as targets and combining Eclipse numerical simulation verification. According to the method, the information in the logging image, the production time sequence data and the structured data can be deeply mined, and the complex association between the gas reservoir productivity and various data can be accurately captured.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploitation, and particularly relates to a method for predicting the productivity of multi-layer combined production in a tight sandstone gas reservoir. Background Art

[0002] At present, in the development process of multi-layer combined production in tight sandstone gas reservoirs, productivity prediction and fracturing parameter optimization are the core links to improve the development efficiency of gas reservoirs, but there are obvious shortcomings in the existing technologies. In terms of productivity prediction, traditional methods are difficult to accurately describe the complex coupling relationships among operating variables (such as pumping pressure and displacement), static parameters (such as reservoir permeability and porosity), and dynamic data (such as production rate and pressure fluctuation), and lack a real-time tracking mechanism for the dynamic changes of gas reservoirs, resulting in the disconnection between the prediction results and the actual productivity, and unable to provide a reliable basis for development decisions. In terms of fracturing parameter optimization, existing means mostly rely on experience or simple models and cannot be dynamically optimized according to the real-time state of the gas reservoir and the accurate productivity prediction results. This not only causes high fracturing operation costs, but also may damage the stable development of the gas reservoir due to the aggravation of interlayer interference, seriously affecting the overall development efficiency. Therefore, the present invention researches and designs a method for predicting the productivity of multi-layer combined production in a tight sandstone gas reservoir. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to overcome the defects of low productivity prediction accuracy and lagging fracturing parameter optimization in the multi-layer combined production of tight sandstone gas reservoirs in the prior art, so as to provide a method for predicting the productivity of multi-layer combined production in a tight sandstone gas reservoir.

[0004] To solve the above problems, the present invention provides a method for predicting the productivity of multi-layer combined production in a tight sandstone gas reservoir, which includes the following steps: S1: Construct a graph data model including operating variables, static parameters, and dynamic data based on the Spearman correlation coefficient, update the node features in real time through Internet of Things sensors, and correct the weights of the adjacency matrix by using a dynamic graph mechanism; S2: Use a CNN network model and an LSTM network model to extract multi-modal data features, input them into an OV-GCN model after PCA dimensionality reduction, and realize graph structure modeling by combining GAT and Transformer, and output the predicted values of productivity distribution, water cut, and interlayer interference; S3: Input the prediction results into an NSGA-II multi-objective optimization model, take productivity maximization, cost minimization, and interference balance as the goals, and verify with Eclipse numerical simulation to generate a Pareto optimal solution set of fracturing parameters.

[0005] Preferably, the step of constructing a graph data model based on the Spearman correlation coefficient includes: performing rank transformation on operating variables, static parameters, and dynamic data, and calculating the rank difference d between node pairs i ; The Spearman correlation coefficient is used to calculate the association strength between nodes to construct an adjacency matrix. The formulas for self-loop correction and normalization processing of the adjacency matrix are as follows: ; ; ; where is the identity matrix, is the transition matrix, A is the original adjacency matrix, A' is the normalized adjacency matrix, adds self-loops to each node.

[0006] Preferably, the real-time update of node features and the correction of adjacency matrix weights through IoT sensors include: Deploy bottom-hole pressure sensors, wellhead flow meters, and water cut meters to collect dynamic data in real time at a frequency of 10 Hz. After filtering outliers through the 3σ criterion by the edge computing node, the data is aggregated into node features at a set time interval; Set the trigger condition for dynamic graph update. When the daily change in water cut ≥ 5%, the fracturing operation is completed, or the static parameters are corrected due to reservoir transformation, start the recalculation of adjacency matrix weights; Set the dynamic update period to 15 minutes, and update the adjacency matrix weights based on the above calculation results. During the update process, retain the historical adjacency matrix versions to support the state backtracking and comparative analysis of the graph structure.

[0007] Preferably, the steps of extracting multi-modal data features using the CNN network model and the LSTM network model include: well logging image feature extraction, production time series data processing, and structured data processing; The steps of inputting into the OV-GCN model after PCA dimensionality reduction include: feature splicing, principal component analysis, and normalization processing.

[0008] Preferably, the steps of combining GAT and Transformer to implement graph structure modeling include: S2.2.1: Connect a graph attention layer (GAT) module after GCN, and connect an 8-head attention mechanism after the graph convolution layer of the OV-GCN model. Perform a linear transformation on the node features through the shared weight matrix and calculate the attention weights between nodes using the LeakyReLU activation function; S2.2.2: Input the node features output by GAT into a 6-layer Transformer encoder, with 8 attention heads set in each layer. Capture the long-range dependencies between nodes through the multi-head self-attention mechanism to generate node embeddings containing global topological information; S2.2.3: Embed the node embeddings output by the Transformer into a fully connected layer, and use the ReLU activation function to predict the production capacity distribution, water cut, and interlayer interference intensity; Meanwhile, introduce the material balance equation constraint to correct the predicted value, and the correction formula is: ; where is a constraint coefficient of 0.1 - 0.3, , are the simulated and measured water saturation, is the original predicted value, is the production capacity predicted value, is the extreme value of water saturation; Characterize the interference intensity through the difference in the predicted production capacity of adjacent small layers, and the formula is: ; where is the thickness of the small layer, 0]is the porosity, which is used to standardize the interlayer production capacity difference.

[0009] Preferably, in the above S2, the data set is divided into a training set, a validation set, and a test set based on historical production data, and the sample labels are the measured production capacity and water cut; Use the multi-task joint loss function to optimize the prediction accuracy and physical consistency of the model, and avoid the model overfitting non-physical laws. The multi-task joint loss formula is: ; where , is the mean square error of production capacity; , is the binary cross-entropy of water cut; , is the mean absolute error of interlayer interference; , is the physical constraint term; is the weight.

[0010] Preferably, the objective function and constraint conditions of the NSGA-II multi-objective optimization model include: The objective function includes: maximizing the production capacity of a single well, minimizing the cost of ineffective fracturing, and balancing interlayer interference; Maximizing the production capacity of a single well is: ; where n is the number of small layers; Minimizing the cost of ineffective fracturing: ; where is the cost of the jth fracturing operation, is the indicator function, is the basic production capacity before fracturing; Interlayer interference balance: ; Among them, is a set of adjacent small layer pairs; The constraint condition is: ; Among them is the fracture length, is the conductivity, is the injection pressure, is the breakdown pressure of the small layer.

[0011] Preferably, the steps of verifying with Eclipse numerical simulation include: Input the productivity distribution predicted by the OV-GCN model , water saturation and fracturing parameters into Eclipse to establish a three-dimensional seepage model with a grid accuracy of 10m×10m×5m, including n small layers; Use the pressure field matching degree and material balance error as simulation verification indicators. The pressure field matching degree is calculated by the root mean square error of the simulated bottom-hole flowing pressure and the measured value , and the threshold is ≤1Mpa; the material balance error is calculated by verifying the difference between the cumulative gas production and the actual value , and the formula is: ; finally, screen the solution set, and only retain the fracturing schemes with δ G ≤3% and the deviation between and ≤1.5% to enter the Pareto solution set.

[0012] Preferably, the strategy for generating the Pareto optimal solution set of fracturing parameters includes: Hybrid optimization algorithm, dynamic weight adjustment, and real-time solution set update.

[0013] Preferably, it further includes S4: Dynamically update the weights of the NSGA-II multi-objective optimization model based on real-time data; The weights of the updated NSGA-II multi-objective optimization model are: Update the OV-GCN model using an incremental training strategy, retain the pre-trained parameters as the initial weights, and only perform gradient backpropagation on the node features corresponding to the real-time data, with an update period ≤10 minutes; at the same time, recalculate the Spearman correlation coefficient and correct the weights of the adjacency matrix to reflect the latest variable associations.

[0014] The method for predicting the productivity of multi-layer combined production in a tight sandstone gas reservoir provided by the present invention has the following beneficial effects: 1. The present invention comprehensively uses various deep learning models such as CNN and LSTM to extract features from multi-modal data, and then combines technologies such as PCA dimensionality reduction, OV-GCN model, GAT, and Transformer to realize graph structure modeling. It can deeply mine the information in logging images, production time-series data, and structured data, accurately capture the complex relationships between gas reservoir productivity and various data, can capture the dynamic changes of the gas reservoir in real time, improve the accuracy of productivity prediction, and timely optimize the fracturing parameters, thereby effectively solving the problems of low productivity prediction accuracy and lagging optimization of fracturing parameters in the prior art, and improving the development efficiency of multi-layer combined production of tight sandstone gas reservoirs; 2. The present invention also inputs the prediction results into the NSGA-II multi-objective optimization model, optimizes with the goals of maximizing productivity, minimizing cost, and balancing interference, and combines Eclipse numerical simulation verification to generate the Pareto optimal solution set of fracturing parameters; the solution set fully considers multiple key factors in gas reservoir development and provides a series of alternative optimal fracturing schemes under different constraint conditions; 3. The present invention also dynamically updates the model weights based on real-time data to form a "data acquisition - prediction - optimization - verification" closed loop, which can timely capture various changes in the gas reservoir production process, such as fluctuations in bottom-hole flowing pressure, production, water cut, etc.; when these data change beyond the threshold, the model will automatically start the online learning process and quickly adjust the prediction and optimization results to ensure that the fracturing parameters always adapt to the latest state of the gas reservoir. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] As Figure 1 shown, the present invention provides a method for predicting the productivity of multi-layer combined production of tight sandstone gas reservoirs, which includes the following steps: S1: Construct a graph data model including operating variables, static parameters, and dynamic data based on the Spearman correlation coefficient, update the node features in real time through Internet of Things sensors, and correct the weights of the adjacency matrix using the dynamic graph mechanism; S2: Use CNN and LSTM to extract the features of multi-modal data, input them into the OV-GCN model after PCA dimensionality reduction, and combine GAT and Transformer to realize graph structure modeling, and output the predicted values of productivity distribution, water cut, and interlayer interference; S3: Input the prediction results into the NSGA-II multi-objective optimization model, with the goals of maximizing productivity, minimizing cost, and balancing interference, and combine Eclipse numerical simulation verification to generate the Pareto optimal solution set of fracturing parameters.

[0017] Specifically, by comprehensively applying various deep learning models such as the CNN network model (Convolutional Neural Network) and LSTM (Long Short-Term Memory) to extract features from multi-modal data, and then combining techniques such as PCA (Principal Component Analysis) dimensionality reduction, the OV-GCN model (Operating Variable-based Graph Convolutional Network), GAT (Graph Attention Network), and the Transformer model to achieve graph structure modeling, it is possible to deeply mine the information in logging images, production time-series data, and structured data, accurately capture the complex relationships between gas reservoir productivity and various types of data. Compared with traditional single-model prediction methods, the accuracy of productivity distribution, water cut, and interlayer interference prediction has been greatly improved; due to the complex geological conditions of tight sandstone gas reservoirs, traditional methods often have difficulty comprehensively considering the impact of various factors on productivity. The present invention constructs a graph data model based on the Spearman correlation coefficient and modifies the weights of the adjacency matrix through a dynamic graph mechanism, which can effectively handle the non-linear relationships between operating variables, static parameters, and dynamic data, and adapt to the gas reservoir characteristics under different geological conditions. Even in gas reservoirs with complex fracture development and large physical property differences, it can accurately predict productivity, broadening the applicable range of productivity prediction methods.

[0018] Specifically, the prediction results are input into the NSGA-II (Non-dominated Sorting Genetic Algorithm II) multi-objective optimization model, and optimized with the goals of maximizing productivity, minimizing cost, and balancing interference. Combined with the numerical simulation verification of Eclipse (Eclipse reservoir numerical simulation software), it is possible to generate the Pareto (Pareto optimal) optimal solution set of fracturing parameters; the solution set fully considers multiple key factors in gas reservoir development and provides a series of alternative optimal fracturing schemes under different constraint conditions; for example, by optimizing parameters such as fracture length, conductivity, and injection volume, the average productivity of a single well can be increased, while effectively reducing ineffective fracturing operations, improving the pertinence and effectiveness of fracturing measures; traditional fracturing parameter design often only focuses on one aspect, such as increasing productivity or reducing cost, while ignoring other factors; the multi-objective optimization model of the present invention can find the best balance point among productivity, cost, and interlayer interference, realizing the sustainable development of gas reservoirs; while pursuing the maximization of productivity, reasonably controlling the fracturing cost, reducing interlayer interference, avoiding irreversible damage to the gas reservoir, and improving the overall development efficiency of the gas reservoir.

[0019] Specifically, through accurate production capacity prediction and multi-objective optimization, it is possible to accurately determine which fracturing operations are effective and which are ineffective, thereby avoiding unnecessary fracturing investments. Under the objective constraint of minimizing the cost of ineffective fracturing, the optimization model will automatically screen out the fracturing plan with the best cost-benefit, reducing the number and scale of ineffective fracturing operations. The present invention can reasonably allocate fracturing resources, such as fracturing fluid, proppant, etc., according to the actual situation and prediction results of the gas reservoir. By optimizing fracturing parameters, resources can be utilized more effectively, avoiding waste of resources. At the same time, the real-time dynamic update mechanism can adjust the fracturing plan in a timely manner according to the changes during the production process of the gas reservoir, further improving the resource allocation efficiency.

[0020] Specifically, based on real-time data to dynamically update the model weights, forming a "data collection - prediction - optimization - verification" closed loop, it can timely capture various changes during the production process of the gas reservoir, such as fluctuations in bottom-hole flowing pressure, production rate, water cut, etc. When these data change beyond the threshold, the model will automatically start the online learning process and quickly adjust the prediction and optimization results to ensure that the fracturing parameters always adapt to the latest state of the gas reservoir. For example, when abnormal situations such as water breakthrough occur in the gas reservoir, the fracturing plan can be adjusted within a short time to reduce the impact on gas reservoir production. The real-time dynamic update mechanism enables the gas reservoir development process to be adjusted in real time according to the actual situation, avoiding production instability caused by changes in gas reservoir conditions. By continuously optimizing fracturing parameters, the stability of the gas reservoir production capacity is maintained, reducing the impact of production rate fluctuations on the economic benefits of the gas field. At the same time, it also reduces the risks during the gas reservoir development process and improves the safety and reliability of gas reservoir development.

[0021] Specifically, the method of the present invention organically combines production capacity prediction, fracturing parameter optimization, and real-time dynamic adjustment to form a complete gas reservoir development solution. Through accurate prediction and optimization, the potential of the gas reservoir can be fully exploited, and the recovery rate of the gas reservoir can be improved. In practical applications, the average recovery rate of the gas reservoir is increased, greatly extending the development life of the gas reservoir, improving the overall development efficiency of the gas reservoir, and realizing the automation and intelligence of the gas reservoir development process. From data collection, processing to model prediction, optimization, and then to real-time dynamic adjustment, the entire process requires no manual intervention, greatly improving the management level and decision-making efficiency of gas reservoir development.

[0022] In some embodiments, the steps of constructing the graph data model based on the Spearman correlation coefficient include: performing rank transformation on operating variables (fracture length, conductivity, injection volume), static parameters (permeability, porosity, water saturation), and dynamic data (real-time pressure, production rate), and calculating the rank difference d between node pairs i ; The Spearman correlation coefficient is used to calculate the association strength between nodes to construct an adjacency matrix. The formula for the Spearman correlation coefficient is: ; where d i is the rank difference between variables i and j, and n is the number of samples. When the absolute value of the correlation coefficient < 0.3, the edge connection is ignored; The formula for self-loop correction and normalization processing of the adjacency matrix is: ; ; ; where is the identity matrix, is the transition matrix, A is the original adjacency matrix, and A' is the normalized adjacency matrix, adds self-loops to each node.

[0023] In some embodiments, the real-time update of node features and the correction of the adjacency matrix weights through the Internet of Things sensors include: Deploy bottom-hole pressure sensors, wellhead flowmeters, and water cut meters to collect dynamic data in real time at a frequency of 10 Hz. After filtering outliers through the 3σ criterion by the edge computing node, the data is aggregated into node features at 1-minute intervals; Set the trigger condition for dynamic graph update. When the daily change in water cut ≥ 5%, the fracturing operation is completed (triggered by the pump injection pressure zero signal), or the static parameters (such as permeability) are corrected due to reservoir transformation, start the recalculation of the adjacency matrix weights; Set the dynamic update period to 15 minutes, and update the adjacency matrix weights based on the above calculation results. During the update process, retain the historical adjacency matrix versions (timestamp index) to support the state backtracking and comparative analysis of the graph structure.

[0024] In some embodiments, the extraction of multi-modal data features using the CNN network model and the LSTM network model includes: well logging image feature extraction, production time series data processing, and structured data processing; Specifically, in feature splicing, 128-dimensional image features, 32-dimensional time series features, and 10-dimensional structured features are spliced into a 170-dimensional original feature vector; Principal component analysis uses the PCA algorithm to reduce the dimension of the 170-dimensional features, retain the principal components with a cumulative variance contribution rate of z95%, and generate a 50-dimensional feature matrix; Normalization processing performs Z-score normalization on the dimension-reduced feature matrix. The formula is: ; where and They are the mean and standard deviation of the training set features respectively, and are input into the OV-GCN model after processing.

[0025] Specifically, the steps of inputting into the OV-GCN model after PCA dimensionality reduction include: feature splicing, principal component analysis, and normalization processing.

[0026] Among them, for well logging image feature extraction, a well logging image with a resolution of 512×512 pixels is used, input into a pre-trained ResNet-50 convolutional neural network, and 2048-dimensional features output by the average pooling layer of the penultimate layer are extracted and reduced to 128 dimensions through max pooling; For production time series data processing, 7-day historical data (1008 time points in the time window) of real-time pressure, production, and water cut are input into a 2-layer LSTM network, with 128 hidden units set in each layer and Dropout (retention rate 0.8) added, and 32-dimensional time series features are output; For structured data processing, fracture parameters (fracture length, conductivity) and static physical property parameters (permeability, porosity, water saturation) are Min-Max normalized to the [0, 1] interval to generate a 10-dimensional structured feature vector.

[0027] In some embodiments, the steps of implementing graph structure modeling by combining GAT and Transformer include: S2.2.1: Connect a graph attention layer (GAT) module after GCN, and connect an 8-head attention mechanism after the graph convolution layer of the OV-GCN model. Through the shared weight matrix perform a linear transformation on the node features, and use the LeakyReLU activation function to calculate the attention weights between nodes. The formula is: ; where is a 128-dimensional attention vector, and its dimension is , and after weighted aggregation of the neighbor node features, 64-dimensional features are output, , represents the set of neighbor nodes of node i, is a trainable weight matrix, , represent the feature vectors of nodes i and j respectively, is the transpose after linear transformation of the feature of node i, is the linear transformation of the feature of node j, is to take the exponential of the activation result, turn negative numbers into positive numbers and amplify the numerical difference to highlight the importance of the association between nodes; It means to repeat the molecular calculation and sum for all neighbors k of node i, which is used for molecular normalization to ensure that the sum of the attention weights of all neighbor nodes is 1; S2.2.2: Input the node features output by GAT into a 6-layer Transformer encoder, with 8 attention heads set in each layer. Through the multi-head self-attention mechanism, capture the long-range dependencies between nodes and generate node embeddings containing global topological information; Input the node embeddings output by Transformer into a fully connected layer, and use the ReLU activation function to predict the production capacity distribution, water cut, and interlayer interference intensity; At the same time, introduce the material balance equation constraint to correct the predicted value, and the correction formula is: ; Among them, is the constraint coefficient of 0.1 - 0.3, , is the simulated and measured water saturation, is the original predicted value, is the production capacity predicted value, is the extreme value of water saturation; Characterize the interference intensity through the difference in the production capacity predicted values of adjacent small layers, and the formula is

[0028] ; Among them is the thickness of the small layer, is the porosity, which is used to standardize the interlayer production capacity difference, , are the corrected production capacity predicted values of the i-th and j-th small layers respectively.

[0029] In some embodiments, in S2, the dataset is divided into a training set (70%), a validation set (20%), and a test set (10%) based on historical production data (5-year cycle), and the sample labels are the measured production capacity and water cut; Use the multi-task joint loss function to optimize the prediction accuracy and physical consistency of the model, and avoid the model overfitting non-physical laws. The multi-task joint loss formula is: ; Among them, , is the mean square error of production capacity; , is the binary cross-entropy of water cut; , is the mean absolute error of interlayer interference; , is the physical constraint term; is the weight.

[0030] In some embodiments, the objective functions and constraint conditions of the NSGA-II multi-objective optimization model include: The objective functions include: maximizing the productivity of a single well, minimizing the cost of ineffective fracturing, and balancing the interlayer interference; Maximizing the productivity of a single well is: ; where n is the number of small layers; Minimizing the cost of ineffective fracturing: ; where is the total number of fracturing operations, is the cost of the j-th fracturing operation, is an indicator function (when the condition in the parentheses holds, I = 1, otherwise I = 0), is the basic productivity before fracturing, is the production change of the j-th unit; Balancing the interlayer interference: ; where is the set of adjacent small layer pairs; The constraint condition is: ; where is the fracture length, is the conductivity, is the injection pressure, is the breakdown pressure of the small layer.

[0031] In some embodiments, the steps of verifying by combining Eclipse numerical simulation include: Input the productivity distribution , water saturation and fracturing parameters predicted by the OV-GCN model into Eclipse to establish a three-dimensional seepage model with a grid accuracy of 10m×10m×5m, including n small layers; Use the pressure field matching degree and material balance error as simulation verification indicators. The pressure field matching degree is calculated by the root mean square error (RMSE) of the simulated bottom hole flowing pressure and the measured value , and the threshold is ≤1 Mpa; the material balance error is calculated by verifying the difference between the cumulative gas production and the actual value , and the formula is: Finally, screen the solution set and only retain the δ G ≤3% and and Fracturing plans with a deviation ≤ 1.5% enter the Pareto solution set.

[0032] In some embodiments, the strategy for generating the Pareto optimal solution set of fracturing parameters includes: Hybrid optimization algorithm, using a hybrid algorithm that combines NSGA-II and particle swarm optimization (PSO). The initial population size is 200, where 70% of the individuals are generated by genetic algorithm (crossover probability 0.8, polynomial mutation distribution index 20), and 30% are generated by PSO (inertia weight linearly decreasing from 0.9 to 0.4); Dynamic weight adjustment, according to the interlayer pressure difference simulated by Eclipse Dynamically adjust the weights of the objective function. The formula is: ; where is the initial value of is the weight of the interlayer interference balance objective, with an initial value of 0.3, is the indicator function; Real-time solution set update. When the daily fluctuation amplitude of the bottom-hole flowing pressure detected by the Internet of Things sensor triggers the online learning update of the OV-GCN model, regenerates the prediction results and starts the NSGA-II optimization. The update period is 1 hour.

[0033] In some embodiments, it further includes S4: Dynamically update the weights of the NSGA-II multi-objective optimization model based on real-time data. The update of the weights of the NSGA-II multi-objective optimization model is as follows: Adopt an incremental training strategy to update the OV-GCN model, retain the pre-trained parameters as the initial weights, and only perform gradient backpropagation on the node features corresponding to the real-time data (such as the permeability nodes associated with pressure changes). The update period ≤ 10 minutes; at the same time, recalculate the Spearman correlation coefficient and correct the weights of the adjacency matrix to reflect the latest variable associations.

[0034] Specifically, it is also updated by real-time data driving: Dynamically collect dynamic data such as bottom-hole flowing pressure, production, and water cut in real time at a frequency of 10 Hz through Internet of Things sensors. After filtering outliers by the 3σ criterion and aggregating at 1-minute intervals through edge computing nodes, trigger model update detection; when it is detected that the real-time data changes exceed the threshold (daily water cut change ≥ 5% or bottom-hole flowing pressure fluctuation ≥ 10%), start the online learning process.

[0035] Specifically, closed-loop verification and policy iteration are also utilized: the updated model prediction results are input into the NSGA-II optimization model, and combined with Eclipse numerical simulation verification (pressure field fitting degree RMSE ≤ 1 MPa, material balance error ≤ 3%) to generate a new Pareto solution set; if it is verified that the prediction error of the interlayer interference intensity > 15%, it will automatically backtrack to the data acquisition link and trigger encrypted sensor monitoring (such as increasing the microseismic monitoring frequency) to form a complete closed loop of "data acquisition → feature update → model prediction → optimization verification → feedback adjustment".

[0036] Taking a certain gas field in Ordos as an example, sensors such as downhole pressure gauges and wellhead flow meters are deployed in 15 wells, and 120 million real-time data records are accumulated; through the method of the present invention, accurate prediction of gas reservoir productivity and effective regulation of fracturing parameters are realized; the injection pressure is controlled within 85% of the fracture pressure, and no formation fracturing accident occurs; the difference rate of interlayer injection volume is reduced from 40% to 15%, and the incidence rate of water channeling is reduced by 60%, significantly improving the gas reservoir development efficiency.

[0037] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention. The above is only the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and variations can be made, and these improvements and variations should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting the production capacity of multi-layer combined production in a tight sandstone gas reservoir, characterized in that, It includes the following steps: S1: Construct a graph data model including operating variables, static parameters, and dynamic data based on the Spearman correlation coefficient. Real-time update the node features through Internet of Things sensors, and use the dynamic graph mechanism to correct the weights of the adjacency matrix; S2: Use the CNN network model and the LSTM network model to extract multi-modal data features. After PCA dimensionality reduction, input them into the OV-GCN model, and combine GAT and Transformer to realize graph structure modeling, and output the predicted values of production capacity distribution, water cut, and interlayer interference; S3: Input the prediction results into the NSGA-II multi-objective optimization model. With the goals of maximizing production capacity, minimizing cost, and balancing interference, combine with Eclipse numerical simulation verification to generate the Pareto optimal solution set of fracturing parameters.

2. The method for predicting the production capacity of multi-layer combined production of tight sandstone gas reservoirs according to claim 1, characterized in that: The steps of constructing a graph data model based on the Spearman correlation coefficient include: performing rank transformation on operating variables, static parameters, and dynamic data, and calculating the rank difference d between node pairs i ; The Spearman correlation coefficient is used to calculate the correlation strength between nodes to construct the adjacency matrix. The self-loop correction and normalization processing formula for the adjacency matrix is: ； ； ； where is the identity matrix, is the transition matrix, A is the original adjacency matrix, and A' is the normalized adjacency matrix, adds self-loops to each node.

3. The method for predicting the production capacity of multi-layer combined production of tight sandstone gas reservoirs according to claim 2, characterized in that: The real-time update of node features and correction of adjacency matrix weights through Internet of Things sensors includes: Deploy bottom-hole pressure sensors, wellhead flow meters, and water cut meters to collect dynamic data in real time at a frequency of 10 Hz. After filtering out outliers through the 3σ criterion by edge computing nodes, aggregate them into node features at a set time interval; Set the dynamic graph update trigger condition. When the daily change in water cut ≥ 5%, the fracturing operation is completed, or the static parameters are corrected due to reservoir transformation, start recalculating the weights of the adjacency matrix; Set the dynamic update period to 15 minutes, and update the weights of the adjacency matrix based on the above calculation results. Retain the historical version of the adjacency matrix during the update process to support the state backtracking and comparative analysis of the graph structure.

4. The method for predicting the production capacity of multi-layer combined production of tight sandstone gas reservoirs according to claim 1, characterized in that: The steps of using the CNN network model and the LSTM network model to extract multi-modal data features include: logging image feature extraction, production time series data processing, and structured data processing; The steps of inputting into the OV-GCN model after PCA dimensionality reduction include: feature splicing, principal component analysis, and normalization processing.

5. The method for predicting the production capacity of multi-layer combined production of tight sandstone gas reservoirs according to claim 4, characterized in that: The steps of combining GAT and Transformer to realize graph structure modeling include: S2.2.1: Connect the Graph Attention Layer (GAT) module after the GCN, and connect the 8-head attention mechanism after the graph convolution layer of the OV-GCN model. Perform a linear transformation on the node features through a shared weight matrix and calculate the attention weights between nodes using the LeakyReLU activation function; S2.2.2: Input the node features output by GAT into a 6-layer Transformer encoder, with 8 attention heads set for each layer, and capture the long-distance dependence relationship between nodes through the multi-head self-attention mechanism to generate node embeddings containing global topological information; S2.2.3: Input the node embeddings output by Transformer into a fully connected layer, and use the ReLU activation function to predict the production capacity distribution, water cut, and interlayer interference intensity; At the same time, introduce the material balance equation constraint to correct the predicted value, and the correction formula is: ； Among them, is a constraint coefficient of 0.1 - 0.3, , is the simulated and measured water saturation, is the original predicted value, is the production capacity predicted value, is the extreme value of water saturation; Characterize the interference intensity through the difference in the predicted production capacity values of adjacent small layers, and the formula is: ； wherein is the small layer thickness, is the porosity, used to standardize the interlayer productivity difference.

6. The method for predicting the multi-layer combined production capacity of a tight sandstone gas reservoir according to claim 5, wherein: In S2, the data set is divided into a training set, a validation set, and a test set based on historical production data, and the sample labels are the measured production capacity and water cut; The multi-task joint loss function is used to optimize the prediction accuracy and physical consistency of the model, and to avoid the model overfitting non-physical laws. The multi-task joint loss formula is: ； Among them, , is the mean square error of production capacity; , is the binary cross-entropy of water cut; , is the mean absolute error of interlayer interference; , is the physical constraint term; is the weight.

7. The method for predicting the multi-layer combined production capacity of a tight sandstone gas reservoir according to claim 1, wherein: The objective function and constraint conditions of the NSGA-II multi-objective optimization model include: The objective function includes: maximizing the single-well production capacity, minimizing the ineffective fracturing cost, and balancing the interlayer interference; Maximizing the single-well production capacity is: ； where n is the number of small layers; Minimizing the ineffective fracturing cost: ； wherein is the cost of the j-th fracturing operation, is an indicator function, basic production capacity before fracturing; Balancing the interlayer interference: ; wherein, is a set of adjacent small layer pairs; The constraint condition is: ； wherein is the fracture length, is the conductivity, is the injection pressure, is the formation breakdown pressure of the sub-layer.

8. The method for predicting the multi-layer combined production capacity of a tight sandstone gas reservoir according to claim 6, wherein: The steps of verifying by combining Eclipse numerical simulation include: The production capacity distribution predicted by the OV-GCN model , water saturation and the fracturing parameters are input into Eclipse to establish a three-dimensional seepage model with a grid accuracy of 10m×10m×5m, including n small layers; Using the pressure field matching degree and material balance error as simulation verification indicators, the pressure field matching degree is calculated by simulating the bottom hole flowing pressure and the measured value of the root mean square error, with a threshold of ≤ 1 Mpa; the material balance error is calculated by verifying the cumulative gas production and the actual value The difference is calculated, and the formula is: ； Finally, the solution set is screened, and only the δ after simulation verification is retained G ≤ 3% and with the deviation ≤ 1.5% of the fracturing plan enters the Pareto solution set.

9. The method for predicting the multi-layer combined production capacity of a tight sandstone gas reservoir according to claim 7, wherein: The strategy for generating the Pareto optimal solution set of fracturing parameters includes: Hybrid optimization algorithm, dynamic weight adjustment, and real-time solution set update.

10. The method for predicting the multi-layer combined production capacity of a tight sandstone gas reservoir according to claim 1, wherein: It further includes S4: dynamically updating the weights of the NSGA-II multi-objective optimization model based on real-time data; Updating the weights of the NSGA-II multi-objective optimization model is: updating the OV-GCN model using an incremental training strategy, retaining the pre-trained parameters as the initial weights, only performing gradient backpropagation on the node features corresponding to the real-time data, and the update period ≤ 10 minutes; at the same time, recalculating the Spearman correlation coefficient and correcting the weights of the adjacency matrix to reflect the latest variable associations.

Images