A method for predicting productivity of multi-layer combined production of a tight sand gas reservoir

Abstract

The present application provides a kind of compact sandstone gas reservoir multi-layer combined production capacity prediction method, method includes the following steps: based on the construction of the graph data model including operating variables, static parameters, dynamic data of spearman correlation coefficient, through Internet of Things sensor real-time update node characteristics, adopt dynamic graph mechanism to revise adjacent matrix weight;S2: utilize CNN network model and LSTM network model to extract multi-modal data features, after PCA dimension reduction, input OV-GCN model, realize graph structure modeling in combination with GAT and Transformer, output capacity distribution, water cut and interlayer interference prediction value;S3: the prediction result is input into NSGA-II multi-objective optimization model, with the maximum capacity, cost minimization, interference balance as target, combined with Eclipse numerical simulation verification, generate fracturing parameter Pareto optimal solution set.According to the present application, the information in the well log image, production time series data and structured data can be deeply mined, and the complex correlation between gas reservoir capacity and various data can be accurately captured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas exploitation, in particular to a multi-layer combined production capacity prediction method for tight sandstone gas reservoirs. BACKGROUND

[0002] At present, in the development process of multi-layer combined production of tight sandstone gas reservoirs, production capacity prediction and fracturing parameter optimization are the core links to improve the development efficiency of gas reservoirs, but the existing technology has obvious shortcomings. From the perspective of production capacity prediction, the traditional method is difficult to accurately depict the complex coupling relationship between operation variables (such as pump injection pressure, displacement), static parameters (such as reservoir permeability, porosity) and dynamic data (such as production, pressure fluctuation), and lacks real-time tracking mechanism for dynamic changes of gas reservoirs, resulting in disconnection between prediction results and actual production capacity, and unable to provide reliable basis for development decision. In terms of fracturing parameter optimization, the existing means mostly rely on experience or simple model, which cannot dynamically optimize according to the real-time state of gas reservoir and accurate production capacity prediction results, which not only causes high cost of fracturing operation, but also may damage the stable development of gas reservoir due to the aggravation of interlayer interference, seriously affecting the overall development efficiency. Therefore, the present application designs a multi-layer combined production capacity prediction method for tight sandstone gas reservoirs. SUMMARY

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

[0004] In order to solve the above problems, the present application provides a multi-layer combined production capacity prediction method for tight sandstone gas reservoirs, which comprises the following steps:

[0005] S1: Based on the construction of a graph data model containing operation variables, static parameters and dynamic data, the node features are updated in real time by Internet of Things sensors, and the adjacency matrix weight is corrected by dynamic graph mechanism;

[0006] S2: Use CNN network model and LSTM network model to extract multi-modal data features, input OV-GCN model after PCA dimension reduction, realize graph structure modeling combined with GAT and Transformer, and output production capacity distribution, water cut and interlayer interference prediction value;

[0007] S3: Input the prediction results into the NSGA-II multi-objective optimization model, take the maximum production capacity, minimum cost and interference balance as the target, combine with Eclipse numerical simulation verification, and generate the Pareto optimal solution set of fracturing parameters.

[0008] Preferably, the step of constructing a graph data model based on the Spearman correlation coefficient comprises: rank conversion of the operating variables, static parameters and dynamic data, calculation of the rank difference d between node pairs i ;

[0009] The Spearman correlation coefficient is used to calculate the correlation strength between nodes to construct an adjacency matrix. The formula for self-loop correction and standardization processing of the adjacency matrix is:

[0010] ;

[0011] ;

[0012] ;

[0013] wherein is the unit matrix, is the excess matrix, A is the original adjacency matrix, A' is the standardized adjacency matrix, is the addition of a self-loop to each node.

[0014] Preferably, the step of updating node features in real time and correcting the adjacency matrix weight through the Internet of Things sensor comprises:

[0015] Deploy the bottom hole pressure sensor, wellhead flowmeter and water cut meter to collect dynamic data in real time at a frequency of 10Hz. After filtering abnormal values by the 3σ criterion through the edge computing node, the node features are aggregated at a set time interval;

[0016] Set the dynamic graph update trigger condition. When the daily change of water cut is greater than or equal to 5%, the fracturing operation is completed, or the static parameters are corrected due to reservoir reconstruction, start the adjacency matrix weight recalculation;

[0017] Set the dynamic update period to 15 minutes. Update the adjacency matrix weight based on the above calculation results. Keep the historical adjacency matrix version during the update process to support state backtracking and comparative analysis of the graph structure.

[0018] Preferably, the step of extracting multi-modal data features using the CNN network model and the LSTM network model comprises: well logging image feature extraction, production time series data processing and structured data processing.

[0019] The step of inputting the PCA dimensionality reduction into the OV-GCN model comprises: feature splicing, principal component analysis and normalization processing.

[0020] Preferably, the step of combining GAT and Transformer to realize graph structure modeling comprises:

[0021] S2.2.1: Access the graph attention layer (GAT) module after the GCN, access the 8-head attention mechanism after the graph convolution layer of the OV-GCN model, and calculate the attention weight between nodes by sharing the weight matrix The node features are linearly transformed, and the LeakyReLU activation function is used to calculate the attention weight between nodes.

[0022] S2.2.2: Input the node features output by GAT into a 6-layer Transformer encoder, set 8 attention heads in each layer, capture long-range dependencies between nodes through multi-head self-attention mechanism, and generate node embeddings containing global topological information.

[0023] S2.2.3: Input the node embeddings output by Transformer into a fully connected layer, and use ReLU activation function to predict the production capacity distribution, water cut and interlayer interference strength.

[0024] At the same time, the predicted value is corrected by introducing the material balance equation constraint, and the correction formula is:

[0025] ;

[0026] Where, 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 prediction value, is the water saturation extreme value.

[0027] The interference strength is represented by the difference between the production capacity prediction values of adjacent small layers, and the formula is:

[0028] ;

[0029] Where is the small layer thickness, is the porosity, which is used to standardize the interlayer production capacity difference.

[0030] Preferably, in S2, the data set is divided into training set, validation set and test set based on historical production data, and the sample label is the measured production capacity and water cut.

[0031] Use a multi-task joint loss function to optimize the prediction accuracy and physical consistency of the model, avoid overfitting of the model to non-physical laws, and the multi-task joint loss formula is:

[0032] ;

[0033] Where, , is the mean square error of production capacity; , is the water cut binary cross-entropy. is an average absolute error of interlayer interference; is a physical constraint term; is a weight.

[0034] Preferably, the objective function and constraint conditions of the NSGA-II multi-objective optimization model include:

[0035] The objective function includes: single well productivity maximization, ineffective fracturing cost minimization, and interlayer interference balance;

[0036] The single well productivity maximization is:

[0037] ;

[0038] wherein n is the number of small layers;

[0039] The ineffective fracturing cost minimization is:

[0040] ;

[0041] wherein is the cost of the jth fracturing operation, is an indicator function, is the base productivity before fracturing;

[0042] The interlayer interference balance is:

[0043] ;

[0044] wherein is a set of adjacent small layer pairs;

[0045] The constraint condition is:

[0046] ;

[0047] wherein is the fracture length, is the conductivity, is the injection pressure, is the small layer fracture pressure.

[0048] Preferably, the step of verifying in combination with the Eclipse numerical simulation includes:

[0049] inputting 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 precision of 10m×10m×5m and containing n small layers;

[0050] The pressure field coincidence degree and the material balance error are used as simulation verification indexes, the pressure field coincidence degree is calculated by taking the root mean square error of the simulated bottom hole pressure and the measured value as the index, the threshold value is ≤1Mpa, the material balance error is calculated by verifying the cumulative gas production and the actual value , and the formula is:

[0051] Finally, the solution set is screened, only the fracturing scheme with δ G ≤3% and and deviation ≤1.5% after simulation verification is retained in the Pareto solution set.

[0052] Preferably, the strategy of generating the Pareto optimal solution set of fracturing parameters comprises:

[0053] A hybrid optimization algorithm, dynamic weight adjustment and real-time solution set updating.

[0054] Preferably, it further comprises S4: dynamically updating the weight of the NSGA-II multi-objective optimization model based on real-time data;

[0055] The updated weight of the NSGA-II multi-objective optimization model is: the OV-GCN model is updated by using the incremental training strategy, the pre-training parameters are retained as the initial weight, only the gradient back propagation is performed on the node features corresponding to the real-time data, and the update period is ≤10 minutes; at the same time, the Spearman correlation coefficient is recalculated, and the adjacent matrix weight is corrected to reflect the latest variable correlation.

[0056] The compact sandstone gas reservoir multi-layer combined production capacity prediction method provided by the application has the following beneficial effects:

[0057] 1. The application can deeply mine the information in the well logging image, production time series data and structured data by comprehensively using various deep learning models such as CNN, LSTM and the like to extract features from multi-modal data, combining PCA dimension reduction, OV-GCN model, GAT and Transformer technology to realize graph structure modeling, accurately capturing the complex correlation between gas reservoir capacity and various data, and can capture the dynamic changes of the gas reservoir in real time, improve the accuracy of the capacity prediction, and timely optimize the fracturing parameters, thereby effectively solving the problems of low capacity prediction accuracy and lagging fracturing parameter optimization in the prior art, improving the development benefit of the compact sandstone gas reservoir multi-layer combined production;

[0058] 2. The present application also optimizes by inputting the prediction results into the NSGA-II multi-objective optimization model to maximize production capacity, minimize cost and balance interference, and verifies with Eclipse numerical simulation to generate a Pareto optimal solution set of fracturing parameters; the solution set fully considers multiple key factors of gas reservoir development and provides a series of optimal fracturing schemes under different constraint conditions for selection;

[0059] 3. The present application also dynamically updates the model weight based on real-time data to form a "data collection-prediction-optimization-verification" closed loop, which can timely capture various changes in the production process of gas reservoirs, such as fluctuations in bottom hole flowing pressure, production, water cut, etc.; when these data change beyond the threshold value, the model will automatically start the online learning process to quickly adjust the prediction and optimization results, ensuring that the fracturing parameters always adapt to the latest state of the gas reservoir. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0061] As shown in Figure 1 , the present application provides a method for predicting the production capacity of multi-layer combined production in tight sandstone gas reservoirs, which includes the following steps:

[0062] S1: Based on the Spearman correlation coefficient, a graph data model containing operation variables, static parameters and dynamic data is constructed, the node features are updated in real time by Internet of Things sensors, and the adjacency matrix weight is corrected by using a dynamic graph mechanism;

[0063] S2: Use CNN and LSTM to extract multi-modal data features, input them into the OV-GCN model after PCA dimensionality reduction, realize graph structure modeling by combining GAT and Transformer, and output the production capacity distribution, water cut and interlayer interference prediction values;

[0064] S3: Input the prediction results into the NSGA-II multi-objective optimization model to maximize production capacity, minimize cost and balance interference, and generate a Pareto optimal solution set of fracturing parameters by combining Eclipse numerical simulation verification.

[0065] Specifically, by comprehensively using CNN network model (Convolutional Neural Network), LSTM (Long Short-Term Memory), and other deep learning models to extract features from multi-modal data, and combining PCA (Principal Component Analysis) dimension reduction, OV-GCN model (Operating Variable-based Graph Convolutional Network), GAT (Graph Attention Network), and Transformer model, the information in the well logging image, production time series data and structured data can be deeply mined, the complex correlation between gas reservoir productivity and various data can be accurately captured, and compared with the prediction method of the traditional single model, the accuracy of the prediction of productivity distribution, water cut and interlayer interference is greatly improved; due to the complex geological conditions of the tight sandstone gas reservoir, the traditional method often has difficulty in comprehensively considering the influence of various factors on the productivity. The graph data model is constructed based on the Spearman correlation coefficient, and the adjacency matrix weight is corrected through the dynamic graph mechanism, which can effectively handle the nonlinear relationship between operating variables, static parameters and dynamic data, and adapt to the characteristics of gas reservoirs under different geological conditions. Even in the gas reservoir with complex fracture development and large physical property difference, the productivity can be accurately predicted, and the application range of the productivity prediction method is widened.

[0066] Specifically, the prediction result is input into the NSGA-II (Non-dominated Sorting Genetic Algorithm II) multi-objective optimization model to optimize the maximum productivity, the minimum cost and the balanced interference, and combined with the Eclipse (Eclipse reservoir numerical simulation software) numerical simulation verification, the Pareto optimal solution set of the fracturing parameters can be generated; the solution set fully considers multiple key factors of gas reservoir development, and provides a series of optimal fracturing schemes under different constraint conditions; for example, by optimizing the parameters such as fracture length, conductivity and water injection amount, the average single well productivity can be improved, and the invalid fracturing operation can be effectively reduced, and the pertinence and effectiveness of the fracturing measures are improved; the traditional fracturing parameter design often only focuses on one aspect, such as improving productivity or reducing cost, and ignores other factors; the multi-objective optimization model can find the best balance point between productivity, cost and interlayer interference, and realize the sustainable development of the gas reservoir; while pursuing the maximum productivity, the fracturing cost is reasonably controlled, the interlayer interference is reduced, the irreversible damage to the gas reservoir is avoided, and the overall development benefit of the gas reservoir is improved.

[0067] Specifically, by accurate production prediction and multi-objective optimization, it can be accurately judged which fracturing operation is effective and which is ineffective, thereby avoiding unnecessary fracturing investment; under the constraint of the goal of minimizing the cost of ineffective fracturing, the optimization model will automatically screen out the fracturing scheme with the best cost-effectiveness, reducing the number and scale of ineffective fracturing operations, and the present application can reasonably allocate fracturing resources such as fracturing fluid and proppant according to the actual situation and prediction results of the gas reservoir; by optimizing the fracturing parameters, the resources can be more effectively utilized, and the waste of resources is avoided; at the same time, the real-time dynamic updating mechanism can timely adjust the fracturing scheme according to the changes in the production process of the gas reservoir, further improving the efficiency of resource allocation.

[0068] Specifically, based on the real-time data dynamic updating model weight, a "data collection-prediction-optimization-verification" closed loop is formed, which can timely capture various changes in the production process of the gas reservoir, such as fluctuations in bottom hole flowing pressure, production, water cut, etc.; when these data change beyond the threshold value, the model will automatically start the online learning process to quickly adjust the prediction and optimization results, ensuring that the fracturing parameters always adapt to the latest state of the gas reservoir; for example, when abnormal conditions such as water channeling occur in the gas reservoir, the fracturing scheme can be adjusted in a short time to reduce the impact on the production of the gas reservoir; the real-time dynamic updating 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 the gas reservoir conditions; by continuously optimizing the fracturing parameters, the stability of the gas reservoir productivity is maintained, and the impact of production fluctuations on the economic benefits of the gas field is reduced; at the same time, the risk in the development process of the gas reservoir is also reduced, improving the safety and reliability of the development of the gas reservoir.

[0069] Specifically, the method of the present application organically combines production 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 tapped, and the recovery of the gas reservoir can be improved; in actual application, the recovery of the gas reservoir is improved on average, greatly prolonging the development life of the gas reservoir and improving the overall development efficiency of the gas reservoir, realizing the automation and intelligentization of the development process of the gas reservoir; from data collection and processing to model prediction, optimization and real-time dynamic adjustment, the whole process does not require manual intervention, greatly improving the management level and decision-making efficiency of the development of the gas reservoir.

[0070] In some embodiments, the step of constructing a graph data model based on the Spearman correlation coefficient comprises: performing rank conversion on the operating variables (fracturing fracture length, conductivity, water injection volume), static parameters (permeability, porosity, water saturation), and dynamic data (real-time pressure, production), calculating the rank difference d i ;

[0071] Spearman correlation coefficient is used to calculate the correlation strength between nodes to construct the adjacency matrix, and the formula of Spearman correlation coefficient is:

[0072] ;

[0073] Wherein, d i is the rank difference of variables i and j, n is the sample number, and the edge connection is ignored when the absolute value of the correlation coefficient is less than 0.3;

[0074] The self-loop correction and standardization processing formula of the adjacency matrix is:

[0075] ;

[0076] ;

[0077] ;

[0078] Wherein is the unit matrix, is the excess matrix, A is the original adjacency matrix, A' is the standardized adjacency matrix, is the self-loop added to each node.

[0079] In some embodiments, the real-time updating of node features and the correction of adjacency matrix weights by the Internet of Things sensor comprises:

[0080] Deploy the bottom hole pressure sensor, wellhead flowmeter, and water cut meter to collect dynamic data in real time at a frequency of 10Hz, filter the abnormal values by the 3σ criterion through the edge computing node, and aggregate the node features at an interval of 1 minute;

[0081] Set the dynamic graph update trigger condition, start the adjacency matrix weight recalculation when the daily change of water cut is greater than or equal to 5%, the fracturing operation is completed (pump injection pressure zero signal trigger), or the static parameter (such as permeability) is corrected due to reservoir reconstruction;

[0082] Set the dynamic update period to 15 minutes, update the adjacency matrix weights based on the above calculation results, and retain the historical adjacency matrix version (time stamp index) during the update process to support state backtracking and comparative analysis of the graph structure.

[0083] In some embodiments, the multi-modal data feature extraction using the CNN network model and the LSTM network model comprises logging image feature extraction, production time series data processing, and structured data processing;

[0084] Specifically, in the 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.

[0085] The principal component analysis adopts a PCA algorithm to reduce the dimension of the 170-dimensional features, retains the principal components with a cumulative variance contribution rate of 95%, and generates a 50-dimensional feature matrix;

[0086] The normalized processing performs Z-score standardization on the reduced dimension feature matrix, and the formula is:

[0087] ;

[0088] Wherein and are the mean and standard deviation of the training set features, and the processed input OV-GCN model.

[0089] Specifically, the step of inputting the PCA reduced dimension into the OV-GCN model includes feature splicing, principal component analysis and normalization processing.

[0090] Wherein, the logging image feature extraction adopts a logging image with a resolution of 512x512 pixels, inputs a pre-trained ResNet-50 convolutional neural network, extracts a 2048-dimensional feature output by the second-to-last layer average pooling layer, and reduces the dimension to 128 by maximum pooling;

[0091] The production time series data processing inputs the 7-day historical data (time window 1008 time points) of real-time pressure, production and water cut into a 2-layer LSTM network, each layer is set to 128 hidden units and adds Dropout (retention rate 0.8), and outputs 32-dimensional time series features;

[0092] The structured data processing performs Min-Max normalization on the fracturing parameters (fracture length, flow conductivity) and static physical parameters (permeability, porosity, water saturation) to the [0, 1] interval, and generates a 10-dimensional structured feature vector.

[0093] In some embodiments, the step of combining GAT with Transformer to realize graph structure modeling includes:

[0094] S2.2.1: Access the graph attention layer (GAT) module after the GCN, access the 8-head attention mechanism after the graph convolution layer of the OV-GCN model, and calculate the attention weight between nodes by sharing the weight matrix The linear transformation is performed on the node features, the LeakyReLU activation function is used to calculate the attention weight between nodes, and the formula is:

[0095] ;

[0096] Wherein is a 128-dimensional attention vector with a dimension of , the neighbor node features are weighted and aggregated to output 64-dimensional features, , set of neighbor nodes of node i, trainable weight matrix, , feature vector of node i and j, respectively, is the transpose of the linear transformation of the features of node i, is the linear transformation of the features of node j, exponentiation of the activation result, conversion of negative numbers to positive numbers and amplification of numerical differences, highlighting the importance of the association between nodes; denotes the summation of the repeated calculation of the numerator for all neighbors k of node i, used for numerator normalization to ensure that the sum of the attention weights of all neighbor nodes is 1;

[0097] S2.2.2: Input the node features output by GAT into a 6-layer Transformer encoder, set 8 attention heads in each layer, capture long-distance dependencies between nodes through multi-head self-attention mechanism, and generate node embeddings containing global topological information;

[0098] Input the node embeddings output by Transformer into a fully connected layer, use ReLU activation function to predict the productivity distribution, water cut and interlayer interference strength;

[0099] At the same time, introduce the material balance equation to constrain and correct the predicted value, and the correction formula is:

[0100] ;

[0101] where, is a constraint coefficient of 0.1-0.3, , is the simulated and measured water saturation, is the original predicted value, is the productivity prediction value, is the water saturation extreme value;

[0102] The interference strength is represented by the difference between the productivity prediction values of adjacent small layers, and the formula is

[0103] ;

[0104] where is the small layer thickness, is the porosity, used to standardize the interlayer productivity difference, , are the corrected productivity prediction values of the i and j small layers, respectively.

[0105] In some embodiments, in the S2, the data set 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 label is the measured production capacity and moisture content;

[0106] The multi-task joint loss function is used to optimize the prediction accuracy and physical consistency of the model, and to avoid overfitting of non-physical laws. The multi-task joint loss formula is:

[0107] ;

[0108] Among them, is the production capacity mean square error; is the moisture content binary cross-entropy; is the interlayer interference average absolute error; is the physical constraint term; is the weight.

[0109] In some embodiments, the objective function and constraint condition of the NSGA-II multi-objective optimization model include:

[0110] The objective function includes: single-well production capacity maximization, invalid fracturing cost minimization, and interlayer interference balance;

[0111] The single-well production capacity maximization is:

[0112] ;

[0113] Among them, n is the number of small layers;

[0114] The invalid fracturing cost minimization is:

[0115] ;

[0116] Among them is the total number of fracturing operations, is the cost of the jth fracturing operation, is an indicator function (I = 1 when the condition in the parentheses is true, otherwise I = 0), is the basic production capacity before fracturing, is the production change of the jth unit;

[0117] The interlayer interference balance is:

[0118] ;

[0119] Among them, is a set of adjacent small layer pairs;

[0120] The constraint condition is: ​

[0121] ;

[0122] wherein is the fracture length, is the conductivity, is the injection pressure, is the fracture pressure.

[0123] In some embodiments, the step of validating the Eclipse numerical simulation comprises:

[0124] The OV-GCN model predicted productivity distribution , water saturation and fracturing parameters are input into Eclipse to establish a three-dimensional seepage model with a grid accuracy of 10m x 10m x 5m, containing n small layers;

[0125] The pressure field coincidence degree and the material balance error are used as simulation validation indicators. The pressure field coincidence degree is calculated by calculating the root mean square error (RMSE) of the simulated bottom hole pressure and the measured value , with a threshold of ≤1Mpa; the material balance error is calculated by verifying the difference between the cumulative gas production and the actual value , with the formula being:

[0126] Finally, the solution set is screened, and only the fracturing scheme with δ G ≤3% and and deviation ≤1.5% after simulation validation is retained into the Pareto solution set.

[0127] In some embodiments, the strategy for generating the Pareto optimal solution set of fracturing parameters comprises:

[0128] Hybrid optimization algorithm, a hybrid algorithm combining NSGA-II and particle swarm optimization (PSO) is used, with an initial population size of 200, of which 70% of the individuals are generated by genetic algorithm (cross probability 0.8, polynomial mutation distribution index 20), and 30% are generated by PSO (inertia weight 0.9-0.4 linearly decreasing);

[0129] Dynamic weight adjustment, the interlayer pressure difference of the Eclipse simulation is used to dynamically adjust the weight of the objective function, with the formula being:

[0130] ;

[0131] wherein is the initial value of , 0.3 for the interlayer interference balance target weight, an initial value 0.3, an indicator function;

[0132] Real-time solution update, when the IoT sensor detects the daily fluctuation amplitude of the bottom-hole flowing pressure When the IoT sensor detects the daily fluctuation amplitude of the bottom-hole flowing pressure

[0133] In some embodiments, S4: dynamically updating the NSGA-II multi-objective optimization model weight based on real-time data is further included. The NSGA-II multi-objective optimization model weight is updated by: updating the OV-GCN model using an incremental training strategy, retaining the pre-training parameters as the initialization weight, and only performing gradient backpropagation on the node features corresponding to the real-time data (such as the permeability node associated with pressure change), and the update period is ≤10 minutes; at the same time, the Spearman correlation coefficient is recalculated, and the adjacency matrix weight is corrected to reflect the latest variable correlation.

[0134] Specifically, the model is further updated in real time by real-time data: the IoT sensor collects dynamic data such as bottom-hole flowing pressure, production, and water cut at a frequency of 10Hz in real time, and after 3σ criterion outlier filtering and 1 minute interval aggregation by the edge computing node, the model update detection is triggered; when the real-time data change is detected to exceed the threshold (water cut daily change ≥5% or bottom-hole flowing pressure fluctuation ≥10%), the online learning process is started.

[0135] Specifically, the updated model prediction result is input into the NSGA-II optimization model, and the new Pareto solution set is generated by combining the Eclipse numerical simulation verification (pressure field coincidence degree RMSE ≤1MPa, material balance error ≤3%); if the verification finds that the interlayer interference strength prediction error is >15%, it is automatically traced back to the data acquisition link, triggering encrypted sensor monitoring (such as increasing the microseismic monitoring frequency), forming a complete closed loop of "data acquisition → feature update → model prediction → optimization verification → feedback adjustment".

[0136] Taking a gas field in Ordos as an example, 120 million real-time data records were obtained by deploying sensors such as bottom-hole pressure gauges and wellhead flowmeters in 15 wells; by the method of the application, the precise prediction of gas reservoir productivity and effective control of fracturing parameters were realized; the injection pressure was controlled within 85% of the fracture pressure, and no formation fracturing accident occurred; the interlayer injection volume difference rate was reduced from 40% to 15%, the water channeling occurrence rate was reduced by 60%, and the gas reservoir development efficiency was significantly improved.

[0137] The above merely describes preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application. The above merely describes preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting deliverability of commingled production in multiple layers of a tight sand gas reservoir, characterized in that, The method comprises the following steps: S1: constructing a graph data model comprising operation variables, static parameters and dynamic data based on a Spearman correlation coefficient, updating node features in real time through Internet of Things sensors, and correcting the weight of the adjacency matrix using a dynamic graph mechanism; S2: extracting multi-modal data features using a CNN network model and an LSTM network model, inputting the features into an OV-GCN model after dimension reduction by PCA, combining GAT and Transformer to realize graph structure modeling, and outputting predicted values of production capacity, water cut and interlayer interference; S3: inputting the predicted results into a NSGA-II multi-objective optimization model, taking maximum production capacity, minimum cost and balanced interference as objectives, verifying the model by Eclipse numerical simulation, and generating a Pareto optimal solution set of fracturing parameters; The step of combining GAT and Transformer to realize graph structure modeling comprises: S2.2.1: Access the graph attention layer (GAT) module after the GCN, access the 8-head attention mechanism after the graph convolution layer of the OV-GCN model, and calculate the attention weight between nodes by sharing the weight matrix Linear transformation is performed on the node features, and LeakyReLU activation function is used to calculate the attention weight between nodes; S2.2.2: inputting the node features output by GAT into a 6-layer Transformer encoder, setting 8 attention heads in each layer, capturing long-distance dependence between nodes through a multi-head self-attention mechanism, and generating node embeddings containing global topological information; S2.2.3: inputting the node embeddings output by Transformer into a fully connected layer, and predicting production capacity distribution, water cut and interlayer interference strength using a ReLU activation function; Meanwhile, the predicted values are corrected by introducing a material balance equation constraint, and the correction formula is: ； wherein, a constraint factor of 0.1-0.3, , is the simulated and measured water saturation, is the original prediction value, is the production capacity prediction value, is the water saturation extreme value; The interference strength is represented by the difference between the predicted values of adjacent sublayers, and the formula is: ； wherein is the small layer thickness, is the porosity, used to normalize the interlayer productivity differences; The objective function and constraint conditions of the NSGA-II multi-objective optimization model comprise: The objective function comprises maximum single-well production capacity, minimum ineffective fracturing cost, and balanced interlayer interference; The maximum single-well production capacity is: ； Where n is the number of sublayers; The minimum ineffective fracturing cost is: ； wherein is the cost of the jth fracturing operation, is an indicator function, pre-fracturing base capacity; The balanced interlayer interference is: ； wherein, is a set of adjacent sublayer pairs; The constraint condition is: ； wherein is the fracture length, is the conductivity, is the injection pressure, is the fracture pressure.

2. The method according to claim 1, characterized in that: The step of constructing the graph data model based on the Spearman correlation coefficient comprises: performing rank conversion on the operation variables, static parameters and dynamic data, and calculating the rank difference d between node pairs i ; The correlation strength between nodes is calculated using a Spearman correlation coefficient to construct an adjacency matrix, and the adjacency matrix is modified with a self-loop and standardized according to the formula: ； ； wherein is the identity matrix, is the excess matrix, A is the original adjacency matrix, A' is the normalized adjacency matrix, adding a self-loop to each node.

3. The method according to claim 2, characterized in that: The node features are updated in real time by Internet of Things sensors and the weight of the adjacency matrix is corrected, comprising: Wellbore pressure sensors, wellhead flow meters and water cut meters are deployed to collect dynamic data in real time at a frequency of 10 Hz, and after filtering abnormal values by 3σ criterion through edge computing nodes, the node features are aggregated at a set time interval; A dynamic graph update trigger condition is set, and when the daily change in water cut is greater than or equal to 5%, fracturing operation is completed, or static parameters are modified due to reservoir reconstruction, the weight of the adjacency matrix is recalculated; The dynamic update period is set to 15 minutes, and the weight of the adjacency matrix is updated based on the above calculation results, and the historical adjacency matrix version is retained during the update process to support state backtracking and comparative analysis of the graph structure.

4. The method according to claim 1, characterized in that: The step of extracting multi-modal data features by using the CNN network model and the LSTM network model comprises well logging image feature extraction, production time series data processing, and structured data processing. The step of inputting the PCA dimension-reduced data into the OV-GCN model comprises feature splicing, principal component analysis, and normalization processing.

5. The method according to claim 4, wherein the method is used for predicting productivity of commingled production of multiple layers in a tight sand gas reservoir. 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 label is measured productivity 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 overfitting of the model to non-physical laws, and the multi-task joint loss formula is: ； wherein, is the capacity mean square error; is the water cut binary cross-entropy; is the inter-layer interference mean absolute error; is the physical constraint term; is the weight.

6. The method according to claim 5, wherein the method is used for predicting productivity of commingled production of multiple layers in a tight sand gas reservoir. The step of combining Eclipse numerical simulation verification comprises: The OV-GCN model predicted productivity distribution , water saturation and fracture parameters into Eclipse to establish a three-dimensional seepage model with a grid accuracy of 10 m x 10 m x 5 m, containing n sublayers; Using pressure field fit and mass balance error as simulation verification indicators, the pressure field fit is calculated by measuring the simulated bottom hole flowing pressure. Compared with measured values The root mean square error is ≤1 MPa; the mass balance error is verified by the cumulative gas production. Compared with actual value The difference is calculated using the following formula: ； Final screening of solution set, only keep δ G ≤ 3% and With Fracturing scheme with deviation ≤ 1.5% enters the Pareto solution set.

7. The method according to claim 1, wherein the method is used for predicting productivity of commingled production of multiple layers in a tight sand gas reservoir. The strategy for generating a Pareto optimal solution set of fracturing parameters comprises: A hybrid optimization algorithm, dynamic weight adjustment, and real-time solution set updating.

8. The method according to claim 1, wherein the method is used for predicting productivity of commingled production of multiple layers in a tight sand gas reservoir. S4: dynamically updating the weight of the NSGA-II multi-objective optimization model based on real-time data is further included. The weight of the NSGA-II multi-objective optimization model is updated by using an incremental training strategy to update the OV-GCN model, retaining the pre-training parameters as the initial weights, performing gradient back propagation only on the node features corresponding to the real-time data, and updating the period ≤ 10 minutes; at the same time, the Spearman correlation coefficient is recalculated, and the adjacency matrix weight is corrected to reflect the latest variable correlation.

Images