A method and system for predicting net pressure in a fracturing wellbore under a data-driven mode

Abstract

The present application relates to a kind of data-driven mode under the method and system for predicting net pressure in fractured well fracture, comprising: collecting oilfield field data, and calculating fracture net pressure;Correlation between well logging parameters and fracture net pressure is calculated and screened, and the well logging parameters after screening are pretreated;Fracture net pressure prediction model is constructed, part of data set is input into fracture net pressure prediction model for training, and the hyperparameters of fracture net pressure prediction model are optimized;Optimized fracture net pressure prediction model is trained using all data sets;When obtaining incremental data, the trained fracture net pressure prediction model is used as submodel, the structure and parameters of submodel are replicated to generate new submodel, and submodel is trained to form submodel for different well sections;The comprehensive prediction value of all submodels is calculated.The present application effectively overcomes the limitation of traditional prediction method, significantly improves the prediction accuracy and model adaptability, and has important significance for oil and gas exploitation engineering.

Description

Technical Field

[0001] This invention relates to a data-driven method and system for predicting net pressure within fractured wells, belonging to the field of oil and gas engineering. Background Technology

[0002] Bottomhole net pressure is a critical parameter in oil and gas extraction, affecting the production efficiency and long-term stability of oil and gas wells. Accurate prediction and control of bottomhole net pressure are essential during the design, construction, and production of oil and gas wells, as it directly determines the flow capacity of oil and gas. By comparing bottomhole net pressure with formation pressure, the degree of reservoir filling and potential production capacity can be assessed. A reasonable bottomhole net pressure ensures smooth flow of oil and gas from the reservoir to the bottom of the well, thereby increasing production. Excessively high bottomhole net pressure can lead to well control problems, such as blowouts and other dangerous situations.

[0003] Bottomhole net pressure is a crucial indicator for evaluating oilfield development effectiveness. Therefore, scientifically predicting bottomhole net pressure using existing data is essential. Currently, the main methods for calculating bottomhole net pressure include well testing, comprehensive evaluation, static pressure gradient, empirical formulas, numerical simulation, analytical solutions, and real-time monitoring technologies. Conventional well testing analysis methods include Horner, MBH, and MDH methods. The Horner curve establishes a logarithmic relationship between bottomhole recovery pressure and shut-in time, using the Horner equation to plot field-measured pressure recovery test data on a semi-logarithmic coordinate system. However, the Horner method requires a long shut-in time to collect sufficient data points, and because it assumes linear pressure recovery, it may not be applicable to all reservoir types. The MDH equation, proposed by Miller CC et al., derives the calculation formula using an approximation of bottomhole recovery pressure. The MDH curve establishes a logarithmic relationship between bottomhole recovery pressure and shut-in time, plots actual field-measured pressure recovery test data on a semi-logarithmic coordinate system, and then performs regression extrapolation on the curve based on actual conditions to determine the formation pressure. The MBH method, proposed by Mathews CS et al., combines the mass balance equation and the Horner equation to derive the MBH function PDBMH(t). PDA ) and dimensionless time t PDAThe MBH method calculates formation pressure by using the relationship curve. However, it requires accurate historical production data; inaccurate data will affect the final calculation results. Lu Detang et al. calculated the formation pressure of a single well based on the pressure distribution integral method, which is applicable to general oil reservoirs. Most oil fields have empirical formulas suitable for their own conditions to predict formation pressure, but these methods lack universality, and changes in reservoir and fluid properties with underground crude oil extraction may render these empirical formulas invalid. Chen Wei used six methods, including the pressure drop curve method, wellhead casing pressure reduction algorithm, fluid mass balance method, and modern production instability analysis method, to calculate the current formation pressure of gas wells and compared the results with the actual pressure measurement values ​​after shut-in. The results showed that most errors were within the allowable error range of the gas field. Among the existing methods, although various prediction models and methods have been proposed to simulate and predict bottomhole net pressure, they each have their limitations. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a data-driven method and system for predicting net pressure within fractured wells. First, relevant oilfield measurement data and fracturing operation data are collected. The correlation between logging parameters and net pressure within the fracture is analyzed using the Pearson correlation coefficient. After data denoising through moving average filtering and normalization, the data is divided into training, validation, and test sets. Next, a Long Short-Term Memory (LSTM) neural network model is constructed, selecting highly correlated parameters as inputs. The Zebra Algorithm (ZOA) is used to optimize key hyperparameters of the model, with normalized weighted sums of mean absolute percentage error (MAPE) and root mean square error (RMSE) used as the overall fitness function. The model is trained using the Adam optimizer, and performance is evaluated through cross-validation. Furthermore, transfer learning is used to fuse historical and incremental data for dynamic model updates. Finally, the weighted sum of sub-model errors is integrated to predict net pressure within the fracture. This invention effectively overcomes the limitations of traditional prediction methods, significantly improves prediction accuracy and model adaptability, and has significant implications for oil and gas extraction engineering.

[0005] Terminology Explanation:

[0006] 1. Pearson correlation coefficient: A statistic used to measure the linear correlation between two variables, with a value ranging from -1 to 1. The closer the absolute value is to 1, the stronger the linear correlation between the variables; the closer the absolute value is to 0, the weaker the linear correlation.

[0007] 2. Adam (Adaptive Moment Estimation): This is an optimization algorithm widely used in deep learning. It aims to train neural network models more efficiently and stably. Adam combines the advantages of RMSprop and Adagrad algorithms, accelerating the training process by adaptively adjusting the learning rate of each parameter.

[0008] The technical solution of the present invention is as follows:

[0009] The first aspect of the present invention provides a method for predicting net pressure within fractured wells in a data-driven mode, comprising:

[0010] Step 1: Collect drilling data, logging parameters, and fracturing operation data from the oilfield site, and calculate the net pressure within the fracture.

[0011] Step 2: Calculate the correlation between logging parameters and net pressure within the fracture and perform screening. Preprocess the screened logging parameters and divide the dataset.

[0012] Step 3: Construct a net intrasuture pressure prediction model. Input a portion of the dataset (70% of the training set and 30% of the validation set) into the net intrasuture pressure prediction model for training. Optimize the hyperparameters of the net intrasuture pressure prediction model to obtain the optimized net intrasuture pressure prediction model.

[0013] Step 4: Train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model.

[0014] Step 5: When acquiring incremental data, use the trained net pressure prediction model within the fracture as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form a sub-model for different well sections; calculate the comprehensive prediction value of all sub-models to predict the net pressure within the fracture in deeper well sections.

[0015] According to a preferred embodiment of the present invention, collecting on-site drilling data, logging parameters, and fracturing operation data from the oilfield, and calculating the net pressure within the fracture; includes:

[0016] Oilfield drilling data includes vertical depth, tubing inner diameter, and tubing length;

[0017] Well logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, wellbore diameter, natural gamma, well depth, well inclination angle, and spontaneous potential;

[0018] Fracturing operation data includes wellhead pressure, hydrostatic column pressure, fracturing fluid friction, near-wellbore friction, and closure stress;

[0019] The hydrostatic pressure is expressed as:

[0020] (1);

[0021] in, Indicates the hydrostatic pressure; Let g represent the density of the liquid column, g represent the acceleration due to gravity, and h represent the height of the liquid column.

[0022] Fracturing fluid friction is expressed as:

[0023] (2);

[0024] in, f Fanning coefficient of friction a and b are constants; This refers to the frictional pressure drop caused by viscous friction during the flow of fracturing fluid. Indicates the fracturing fluid flow rate. This refers to the density of the fracturing fluid. L This refers to the length of the tubing. D p The diameter of the pipeline; N Re,gen It is the generalized Reynolds number;

[0025] Near-wellbore friction is expressed as:

[0026] (3);

[0027] in, p nw For near-wellbore friction; For proppant density; q For displacement; C d For flow coefficient; N perf Number of holes; D perf0 This is the original aperture diameter;

[0028] The net pressure inside the joint is calculated as follows:

[0029] (4);

[0030] in, This refers to the net pressure within the seam. It is the closure stress, approximately equal to the minimum horizontal principal stress; Indicates wellhead pressure, This indicates the hydrostatic pressure.

[0031] According to a preferred embodiment of the present invention, the correlation between logging parameters and net pressure within the fracture is calculated and filtered; the filtered logging parameters are preprocessed and the dataset is divided; including:

[0032] To improve the accuracy of model predictions, it is necessary to select characteristic data and eliminate the influence of irrelevant or weakly correlated parameters. The Pearson correlation coefficient is used to calculate the correlation between logging parameters at different depths and the net pressure within the fracture at the corresponding depth. The calculation formula is as follows:

[0033] (5);

[0034] in, Pearson correlation coefficient; Let X be the covariance of Y; and They are respectively X and Y Standard deviation; X For logging parameters, Y This represents the net pressure within the fracture corresponding to the logging parameters. n The sample size represents the data samples of logging parameters and corresponding net pressure within the fracture at different depths. , For the i-th sample X and Y Observed values; for X The average value; for Y The average value;

[0035] Well logging parameters were selected based on the absolute value of the Pearson correlation coefficient, retaining those with strong correlation (| |≥0.7) and moderate correlation (0.3≤| |<0.7) parameter, remove weakly correlated parameters (| |<0.3) parameters; according to the calculation results, P-wave transit time, S-wave transit time, and rock density are strongly correlated; compensated neutron and wellbore diameter are moderately correlated; natural gamma, well depth, well inclination angle, and spontaneous potential are weakly correlated;

[0036] The filtered logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, and wellbore diameter;

[0037] To improve data quality, a moving average filter is used to filter and reduce noise in the selected logging parameters. The moving average filter uses the average value of the data to replace the original data through moving sampling, thereby reducing the noise in the original data, as shown in the following formula:

[0038] (6);

[0039] in, X k These are the logging parameter values ​​after moving average filtering; L1 The width of the sliding window; k For the first k The selected logging parameters; m The sample size of the sliding window. This represents the mk-th logging parameter;

[0040] Normalization is performed to normalize the well logging parameter values ​​after moving average filtering to between 0 and 1, in order to eliminate dimensional differences between different well logging parameters; the formula is:

[0041] (7);

[0042] in, These are the normalized logging parameter values; These are the well logging parameter values ​​after moving average filtering. and These represent the maximum and minimum values ​​of the logging parameters in the filtered data, respectively.

[0043] The well was divided into a training set, a validation set, and a test set. The training set consisted of normalized logging parameters and net fracture pressure (i.e., net fracture pressure values ​​obtained through comprehensive interpretation of logging data) for the 4839-6563 m well section, which served as the target variable for model training.

[0044] 30% of the data in the training set was randomly selected as the validation set, and the test set consisted of logging parameters and net pressure within the fracture (i.e., net pressure values ​​within the fracture obtained through comprehensive interpretation of logging data) for different well sections (6563 ~ 6994 m) to prepare for subsequent model training and testing.

[0045] According to a preferred embodiment of the present invention, a net intrasuture pressure prediction model is constructed by inputting a portion of the dataset (70% of the training set and 30% of the validation set) into the net intrasuture pressure prediction model for training, and optimizing the hyperparameters of the net intrasuture pressure prediction model to obtain an optimized net intrasuture pressure prediction model; including:

[0046] The intracellular net pressure prediction model is a Long Short-Term Memory (LSTM) neural network. The recurrent unit structure of the LSTM model mainly includes four processes: forgetting phase, selective memory phase, update phase, and output phase. The forgetting phase is defined as the selective forgetting of output data, determining the information to be discarded from the cell state. The calculation formula is as follows:

[0047] (8);

[0048] The selective memory phase selectively identifies new information stored in the cellular state; the calculation formula can be expressed as:

[0049] (9);

[0050] (10);

[0051] The update phase determines and updates the current cell state by combining the forgetting gate, the cell memory value from the previous layer, and the cell memory candidate value. The calculation formula can be expressed as:

[0052] (11);

[0053] The output phase determines the output value of the current state, calculated using the following formula:

[0054] (12);

[0055] (13);

[0056] In the formula: F t for t The state matrix of the forget gate at any time; for Sigmoid function; h t−1 This is the state matrix of the hidden layer at the previous time step; W F , W I , W C , W o They are respectively F t , I t , , o t The weight matrix; b F , b I , b C , b o They are respectively F t , I t , , o t The bias matrix; X t for t The input vector at time step; I t Let be the state matrix of the input gate at time t; for t The candidate value matrix at time step; C t Let be the cell's state matrix after time t; C t−1 for t The state matrix of the cell at time −1; o t for t Output the state matrix of the gate at each time step; h t for t The state matrix of the hidden layer at each time step;

[0057] An LSTM model was constructed based on selected logging parameters (P-wave transit time, S-wave transit time, rock density, compensated neutrons, and natural gamma), and key hyperparameters were optimized using the Zebra Algorithm (ZOA).

[0058] A portion of the training set (70% of the training set) and a portion of the validation set (30% of the validation set) from the dataset are input into the LSTM model for training. During training, the mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE) are normalized and then weighted and summed to form the overall fitness function. The weights of MAPE, RMSE, MAE, and MSE are 0.4, 0.3, 0.2, and 0.1, respectively, as shown in the following formula:

[0059] (14);

[0060] in, The comprehensive fitness function is represented by MAPE, the mean absolute percentage error is represented by RMSE, the root mean square error is represented by MAE, and the mean square error is represented by MSE.

[0061] The fitness function is used to evaluate the quality of the hyperparameter combination of the LSTM model. The smaller the value, the smaller the prediction error of the model. The Zebra algorithm finds the optimal hyperparameter combination by optimizing the fitness function to improve the prediction accuracy of the model.

[0062] The hyperparameters of the LSTM (Long Short-Term Memory Neural Network) model are optimized using the Zebra Algorithm (ZOA) to find the optimal combination of hyperparameters that minimizes the fitness function of the validation set in a subset of the dataset, thus obtaining the optimized intra-fractal net pressure prediction model.

[0063] Hyperparameters include the number of neurons in the hidden layer, learning rate, number of iterations, batch size, and dropout ratio;

[0064] Set the zebra population size (to 20). The zebra population represents the set of individuals participating in hyperparameter optimization. Each individual corresponds to a set of hyperparameter combinations. Set the search space boundary, i.e., the range of hyperparameter values ​​(e.g., number of hidden layer neurons ranges from 10 to 200, learning rate ranges from 0.0001 to 0.1, number of iterations ranges from 50 to 500, batch size ranges from 16 to 128, and dropout ratio ranges from 0.1 to 0.5). Randomly initialize the zebra population (e.g., hidden layer neurons = 50, learning rate = 0.001). Configure an LSTM model for each zebra individual. Train the LSTM model weights on the training set (partial dataset) and calculate the Fitness (i.e., the comprehensive fitness function) as the fitness value on the validation set (partial dataset).

[0065] The zebra with the lowest fitness is selected as the current optimal hyperparameter. The zebra population is then guided to update the hyperparameter combination through foraging behavior (moving towards the leader position) and defensive behavior (random perturbation). This iteration is repeated (up to 100 times) until the fitness converges or the termination condition is met. The optimal hyperparameter combination is then output (number of hidden layer neurons = 128, learning rate = 0.0032, number of iterations = 320, batch size = 64, dropout ratio = 0.23), resulting in the optimized intra-suture net pressure prediction model.

[0066] According to a preferred embodiment of the present invention, the optimized intra-suture net pressure prediction model is trained using the entire dataset to obtain a trained intra-suture net pressure prediction model; including:

[0067] During training, the model continuously adjusts its weights based on the input P-wave transit time, S-wave transit time, rock density, compensated neutrons, and wellbore parameters, and learns the relationship between these parameters and the net pressure within the fracture.

[0068] The entire dataset is input into the optimized intra-fracture net pressure prediction model for training. During training, the Adam optimizer is used to minimize the loss function (such as mean square error MSE) to learn the relationship between logging parameters in the dataset and intra-fracture net pressure.

[0069] The training process is monitored using a validation set (e.g., cross-validation, early stopping) and the intrasuture net pressure prediction model is evaluated. The training set is further divided into multiple subsets using cross-validation. After multiple training and validation cycles, the average index is used to evaluate the intrasuture net pressure prediction model. The predicted and actual intrasuture net pressure data are compared, and the performance of the intrasuture net pressure prediction model is evaluated based on any one of the following indices: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE), as well as the comprehensive fitness function (Formula 14). The training parameters (such as learning rate and number of iterations, which are training parameters, not structural hyperparameters) are adjusted until the intrasuture net pressure prediction model reaches its optimal performance, thus obtaining the trained intrasuture net pressure prediction model.

[0070] For example, the training set can be divided into 5 subsets. Each time, 4 subsets are selected as sub-training sets for model training, and the remaining 1 subset is used for validation. This process is repeated 5 times, and the average of the 5 validation results is taken as the final validation metric. In this way, the performance of the model can be evaluated more comprehensively, avoiding evaluation bias caused by the randomness of the division of the training and validation sets.

[0071] According to a preferred embodiment of the present invention, when acquiring incremental data, the trained intra-fracture net pressure prediction model is used as a sub-model, the structure and parameters of the sub-model are copied to generate a new sub-model, and trained to form sub-models for different well sections; the comprehensive prediction value of all sub-models is calculated to predict the intra-fracture net pressure in deeper well sections; including:

[0072] The trained intra-suture net pressure prediction model was used as a sub-model ZOA-LSTM. 0 ;

[0073] When acquiring incremental data ΔX 1 When the logging parameters and corresponding net pressure within the fracture are set for the new well section (such as the logging parameters and net pressure within the fracture in the 6563~6994m well section or deeper well sections), the ZOA-LSTM sub-model is replicated. 0 Based on the structure and parameters, a new sub-model ZOA-LSTM is generated. 1 Freezing sub-model ZOA-LSTM 1 ZOA-LSTM sub-model 0 Shared core layer weights (such as LSTM hidden layers and feature embedding layers) only allow the weight matrix and bias vector parameters of the output layer to be trained;

[0074] Utilizing incremental data ΔX 1 Training sub-model ZOA-LSTM 1 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 1 ;

[0075] Sub-model ZOA-LSTM 0 With sub-model ZOA-LSTM 1 The output, i.e. the prediction result, is fused through the transfer sharing layer, as shown below:

[0076] (15);

[0077] in, To migrate the shared layer output, F For activation functions; For the original sub-model (such as sub-model ZOA-LSTM) 0 The output of ) For the original sub-model (such as sub-model ZOA-LSTM) 0 The weight matrix; For the generation of new sub-models (such as sub-model ZOA-LSTM) 1 The output of ) For the generation of new sub-models (such as sub-model ZOA-LSTM) 1 The weight matrix;

[0078] When acquiring incremental data ΔX 2 At the same time, the sub-model ZOA-LSTM is used in the same way as described above. 1 Perform replication to generate the sub-model ZOA-LSTM 2 Freeze the weights of core layers (such as LSTM hidden layers and feature embedding layers) and use incremental data. ΔX 2 Training sub-model ZOA-LSTM 2 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 2 ;

[0079] Similarly, the sub-model ZOA-LSTM 1 With sub-model ZOA-LSTM 2 The output, i.e. the prediction result, is fused through the migration sharing layer; this process is repeated to accumulate and generate a set of sub-models, which utilizes the adaptability of multi-well data to improve the prediction generalization.

[0080] Finally, all sub-models are integrated using error weighting, based on the error weights of each sub-model (ZOA-LSTM). 0 ZOA-LSTM 1 The prediction error of each sub-model (e.g., [model name]) is assigned a weight based on its magnitude in the corresponding well section, and the weighted sum is used to obtain the comprehensive predicted value of net pressure within the fracture in deeper well sections. The core logic is that the smaller the error of the sub-model, the greater its contribution to the final prediction; as shown below:

[0081] (16);

[0082] (17);

[0083] in, Let i be the weight of the i-th sub-model. The error value (such as MAE or other indicators) of the i-th sub-model on the corresponding test set is used to calculate the weights. All weights are normalized and then combined to 1. The predicted value of the i-th sub-model for the target well section (e.g., ZOA-LSTM) 1 (Prediction for 6563~6994 m). This represents the final integrated prediction of the net pressure within the fracture in the target well section, i.e., the output after fusing the prediction results of all sub-models;

[0084] When applying the method, the logging parameters of the target well section are input into each sub-model, and the prediction results are output after weighted fusion.

[0085] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a data-driven method for predicting net pressure within fractured wells.

[0086] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a data-driven method for predicting net pressure within fractured wells.

[0087] A second aspect of the present invention provides a data-driven system for predicting net pressure within fractured wells, comprising:

[0088] The data collection module is configured to collect drilling data, logging parameters, and fracturing operation data from the oilfield, and to calculate the net pressure within the fracture.

[0089] The preprocessing module is configured to: calculate the correlation between logging parameters and net pressure within the fracture and filter them; preprocess the filtered logging parameters and divide them into datasets.

[0090] The model optimization module is configured to: construct an intrasuture net pressure prediction model, input a portion of the dataset into the intrasuture net pressure prediction model for training, and optimize the hyperparameters of the intrasuture net pressure prediction model to obtain an optimized intrasuture net pressure prediction model.

[0091] The model training module is configured to train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model.

[0092] The model prediction module is configured to: when incremental data is acquired, use the trained net pressure prediction model within the fracture as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form a sub-model for different well sections; calculate the comprehensive prediction value of all sub-models to predict the net pressure within the fracture in deeper well sections.

[0093] The beneficial effects of this invention are as follows:

[0094] 1. Compared with the prior art, the present invention first collects relevant data, filters parameters through Pearson correlation coefficient analysis, and divides the dataset after moving average filtering for noise reduction and normalization; constructs an LSTM model, optimizes its hyperparameters using the Zebra algorithm, and uses a normalized weighted sum of multiple error indices as the comprehensive fitness function; and uses the Adam optimizer and cross-validation method to improve model performance during training.

[0095] 2. The model can utilize transfer learning to integrate historical and incremental data, achieving dynamic updates. In terms of prediction performance, this invention, through the aforementioned advantages, effectively reduces prediction errors and improves accuracy and reliability. It better meets the stringent requirements for bottom hole net pressure prediction in oil and gas extraction, providing stronger technical support for oil and gas extraction operations. This contributes to improving oil and gas extraction efficiency and safety, demonstrating significant practical value and application prospects. Attached Figure Description

[0096] Figure 1 This is a flowchart of a data-driven method for predicting net pressure within fractured wells according to the present invention.

[0097] Figure 2 This is a diagram of the LSTM cyclic cell structure.

[0098] Figure 3 This is a schematic diagram of the model update and feature fusion of the present invention;

[0099] Figure 4 A comparison chart showing the fitting of actual and predicted net pressure values ​​in the shallow well section 4839-6563m;

[0100] Figure 5 A comparison chart showing the fitting of actual and predicted net pressure values ​​in the middle well section 6563-6994m;

[0101] Figure 6 This is a comparison chart showing the fitting of actual and predicted net pressure values ​​in the deep well section 6994-7500m. Detailed Implementation

[0102] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0103] Example 1

[0104] A data-driven method for predicting net pressure within fractured wells, such as Figure 1 As shown, it includes:

[0105] Step 1: Collect drilling data, logging parameters, and fracturing operation data from the oilfield site, and calculate the net pressure within the fracture.

[0106] Step 2: Calculate the correlation between logging parameters and net pressure within the fracture and perform screening. Preprocess the screened logging parameters and divide them into datasets.

[0107] Step 3: Construct a net intrasuture pressure prediction model. Input a portion of the dataset (70% of the training set and 30% of the validation set) into the net intrasuture pressure prediction model for training. Optimize the hyperparameters of the net intrasuture pressure prediction model to obtain the optimized net intrasuture pressure prediction model.

[0108] Step 4: Train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model.

[0109] Step 5: When acquiring incremental data, use the trained intra-fracture net pressure prediction model as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form a sub-model for different well sections; calculate the comprehensive prediction value of all sub-models to predict the intra-fracture net pressure in deeper well sections.

[0110] Example 2

[0111] The difference between the data-driven method for predicting net pressure within fractured wells as described in Example 1 and the method described in Example 1 is as follows:

[0112] Collect on-site drilling data, logging parameters, and fracturing operation data from the oilfield, and calculate the net pressure within the fracture; including:

[0113] Oilfield drilling data includes vertical depth, tubing inner diameter, and tubing length;

[0114] Well logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, wellbore diameter, natural gamma, well depth, well inclination angle, and spontaneous potential;

[0115] Fracturing operation data includes wellhead pressure, hydrostatic column pressure, fracturing fluid friction, near-wellbore friction, and closure stress;

[0116] The hydrostatic pressure is expressed as:

[0117] (1);

[0118] in, Indicates the hydrostatic pressure; Let g represent the density of the liquid column, g represent the acceleration due to gravity, and h represent the height of the liquid column.

[0119] Fracturing fluid friction is expressed as:

[0120] (2);

[0121] in, f Fanning coefficient of friction a and b are constants; This refers to the frictional pressure drop caused by viscous friction during the flow of fracturing fluid. Indicates the fracturing fluid flow rate. This refers to the density of the fracturing fluid. L This refers to the length of the tubing. D p The diameter of the pipeline; N Re,gen It is the generalized Reynolds number;

[0122] Near-wellbore friction is expressed as:

[0123] (3);

[0124] in, p nw For near-wellbore friction; For proppant density; q For displacement; C d For flow coefficient; N perf Number of holes; D perf0 This is the original aperture diameter;

[0125] The net pressure inside the joint is calculated as follows:

[0126] (4);

[0127] in, This refers to the net pressure within the seam. It is the closure stress, approximately equal to the minimum horizontal principal stress; Indicates wellhead pressure, This indicates the hydrostatic pressure.

[0128] The correlation between well logging parameters and net pressure within the fracture was calculated and filtered. The filtered well logging parameters were preprocessed and divided into datasets, including:

[0129] To improve the accuracy of model predictions, it is necessary to select characteristic data and eliminate the influence of irrelevant or weakly correlated parameters. The Pearson correlation coefficient is used to calculate the correlation between logging parameters at different depths and the net pressure within the fracture at the corresponding depth. The calculation formula is as follows:

[0130] (5);

[0131] in, Pearson correlation coefficient; Let X be the covariance of Y; and They are respectively X and Y Standard deviation; X For logging parameters, Y This represents the net pressure within the fracture corresponding to the logging parameters. n The sample size represents the data samples of logging parameters and corresponding net pressure within the fracture at different depths. , For the i-th sample X and Y Observed values; for XThe average value; for Y The average value;

[0132] Well logging parameters were selected based on the absolute value of the Pearson correlation coefficient, retaining those with strong correlation (| |≥0.7) and moderate correlation (0.3≤| |<0.7) parameter, remove weakly correlated parameters (| |<0.3) parameters; according to the calculation results, P-wave transit time, S-wave transit time, and rock density are strongly correlated; compensated neutron and wellbore diameter are moderately correlated; natural gamma, well depth, well inclination angle, and spontaneous potential are weakly correlated;

[0133] The filtered logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, and wellbore diameter;

[0134] To improve data quality, a moving average filter is used to filter and reduce noise in the selected logging parameters. The moving average filter uses the average value of the data to replace the original data through moving sampling, thereby reducing the noise in the original data, as shown in the following formula:

[0135] (6);

[0136] in, X k These are the logging parameter values ​​after moving average filtering; L1 The width of the sliding window; k For the first k The selected logging parameters; m The sample size of the sliding window. This represents the mk-th logging parameter;

[0137] Normalization is performed to normalize the well logging parameter values ​​after moving average filtering to between 0 and 1, in order to eliminate dimensional differences between different well logging parameters; the formula is:

[0138] (7);

[0139] in, These are the normalized logging parameter values; These are the well logging parameter values ​​after moving average filtering. and These represent the maximum and minimum values ​​of the logging parameters in the filtered data, respectively.

[0140] The well was divided into a training set, a validation set, and a test set. The training set consisted of normalized logging parameters and net fracture pressure (i.e., net fracture pressure values ​​obtained through comprehensive interpretation of logging data) for the 4839-6563 m well section, which served as the target variable for model training.

[0141] 30% of the data in the training set was randomly selected as the validation set, and the test set consisted of logging parameters and net pressure within the fracture (i.e., net pressure values ​​within the fracture obtained through comprehensive interpretation of logging data) for different well sections (6563 ~ 6994 m) to prepare for subsequent model training and testing.

[0142] A net intrasuture pressure prediction model was constructed. A portion of the dataset (70% of the training set and 30% of the validation set) was input into the model for training. The hyperparameters of the model were then optimized to obtain the optimized net intrasuture pressure prediction model. This included:

[0143] The intracellular net pressure prediction model is a Long Short-Term Memory (LSTM) neural network. The recurrent unit structure of the LSTM model mainly includes four processes: forgetting phase, selective memory phase, update phase, and output phase. The forgetting phase is defined as the selective forgetting of output data, determining the information to be discarded from the cell state. Figure 2 As shown, the calculation formula is:

[0144] (8);

[0145] The selective memory phase selectively identifies new information stored in the cellular state; the calculation formula can be expressed as:

[0146] (9);

[0147] (10);

[0148] The update phase determines and updates the current cell state by combining the forgetting gate, the cell memory value from the previous layer, and the cell memory candidate value. The calculation formula can be expressed as:

[0149] (11);

[0150] The output phase determines the output value of the current state, calculated using the following formula:

[0151] (12);

[0152] (13);

[0153] In the formula: F t for t The state matrix of the forget gate at any time; for Sigmoid function; h t−1 This is the state matrix of the hidden layer at the previous time step; W F , WI , W C , W o They are respectively F t , I t , , o t The weight matrix; b F , b I , b C , b o They are respectively F t , I t , , o t The bias matrix; X t for t The input vector at time step; I t Let be the state matrix of the input gate at time t; for t The candidate value matrix at time step; C t Let be the cell's state matrix after time t; C t−1 for t The state matrix of the cell at time −1; o t for t Output the state matrix of the gate at each time step; h t for t The state matrix of the hidden layer at each time step;

[0154] An LSTM model was constructed based on selected logging parameters (P-wave transit time, S-wave transit time, rock density, compensated neutrons, and natural gamma), and key hyperparameters were optimized using the Zebra Algorithm (ZOA).

[0155] A portion of the training set (70% of the training set) and a portion of the validation set (30% of the validation set) from the dataset are input into the LSTM model for training. During training, the mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE) are normalized and then weighted and summed to form the overall fitness function. The weights of MAPE, RMSE, MAE, and MSE are 0.4, 0.3, 0.2, and 0.1, respectively, as shown in the following formula:

[0156] (14);

[0157] in, The comprehensive fitness function is represented by MAPE, the mean absolute percentage error is represented by RMSE, the root mean square error is represented by MAE, and the mean square error is represented by MSE.

[0158] The fitness function is used to evaluate the quality of the hyperparameter combination of the LSTM model. The smaller the value, the smaller the prediction error of the model. The Zebra algorithm finds the optimal hyperparameter combination by optimizing the fitness function to improve the prediction accuracy of the model.

[0159] The hyperparameters of the LSTM (Long Short-Term Memory Neural Network) model are optimized using the Zebra Algorithm (ZOA) to find the optimal combination of hyperparameters that minimizes the fitness function of the validation set in a subset of the dataset, thus obtaining the optimized intra-fractal net pressure prediction model.

[0160] Hyperparameters include the number of neurons in the hidden layer, learning rate, number of iterations, batch size, and dropout ratio;

[0161] Set the zebra population size (to 20). The zebra population represents the set of individuals participating in hyperparameter optimization. Each individual corresponds to a set of hyperparameter combinations. Set the search space boundary, i.e., the range of hyperparameter values ​​(e.g., number of hidden layer neurons ranges from 10 to 200, learning rate ranges from 0.0001 to 0.1, number of iterations ranges from 50 to 500, batch size ranges from 16 to 128, and dropout ratio ranges from 0.1 to 0.5). Randomly initialize the zebra population (e.g., hidden layer neurons = 50, learning rate = 0.001). Configure an LSTM model for each zebra individual. Train the LSTM model weights on the training set (partial dataset) and calculate the Fitness (i.e., the comprehensive fitness function) as the fitness value on the validation set (partial dataset).

[0162] The zebra with the lowest fitness is selected as the current optimal hyperparameter. The zebra population is then guided to update the hyperparameter combination through foraging behavior (moving towards the leader position) and defensive behavior (random perturbation). This iteration is repeated (up to 100 times) until the fitness converges or the termination condition is met. The optimal hyperparameter combination is then output (number of hidden layer neurons = 128, learning rate = 0.0032, number of iterations = 320, batch size = 64, dropout ratio = 0.23), resulting in the optimized intra-suture net pressure prediction model.

[0163] The optimized intra-suture net pressure prediction model was trained using the entire dataset, resulting in a well-trained intra-suture net pressure prediction model; including:

[0164] During training, the model continuously adjusts its weights based on the input P-wave transit time, S-wave transit time, rock density, compensated neutrons, and wellbore parameters, and learns the relationship between these parameters and the net pressure within the fracture.

[0165] The entire dataset was input into the optimized intra-fracture net pressure prediction model for training. During training, the Adam optimizer was used to minimize the loss function (mean square error MSE) to learn the relationship between logging parameters in the dataset and intra-fracture net pressure.

[0166] The training process is monitored using a validation set (e.g., cross-validation, early stopping) and the intrasuture net pressure prediction model is evaluated. The training set is further divided into multiple subsets using cross-validation. After multiple training and validation cycles, the average index is used to evaluate the intrasuture net pressure prediction model. The predicted and actual intrasuture net pressure data are compared, and the performance of the intrasuture net pressure prediction model is evaluated based on any one of the following indices: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE), as well as the comprehensive fitness function (Formula 14). The training parameters (such as learning rate and number of iterations, which are training parameters, not structural hyperparameters) are adjusted until the intrasuture net pressure prediction model reaches its optimal performance, thus obtaining the trained intrasuture net pressure prediction model.

[0167] For example, the training set can be divided into 5 subsets. Each time, 4 subsets are selected as sub-training sets for model training, and the remaining 1 subset is used for validation. This process is repeated 5 times, and the average of the 5 validation results is taken as the final validation metric. In this way, the performance of the model can be evaluated more comprehensively, avoiding evaluation bias caused by the randomness of the division of the training and validation sets.

[0168] When acquiring incremental data, the trained intra-fracture net pressure prediction model is used as a sub-model. The structure and parameters of this sub-model are copied to generate a new sub-model, which is then trained to form sub-models for different well sections. The combined prediction value of all sub-models is calculated to predict the intra-fracture net pressure in deeper well sections. For example... Figure 3 As shown, it includes:

[0169] The trained intra-suture net pressure prediction model was used as a sub-model ZOA-LSTM. 0 ;

[0170] When acquiring incremental data ΔX 1 When the logging parameters and corresponding net pressure within the fracture are set for the new well section (such as the logging parameters and net pressure within the fracture in the 6563~6994m well section or deeper well sections), the ZOA-LSTM sub-model is replicated. 0 Based on the structure and parameters, a new sub-model ZOA-LSTM is generated. 1Freezing sub-model ZOA-LSTM 1 ZOA-LSTM sub-model 0 Shared core layer weights (such as LSTM hidden layers and feature embedding layers) only allow the weight matrix and bias vector parameters of the output layer to be trained;

[0171] Utilizing incremental data ΔX 1 Training sub-model ZOA-LSTM 1 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 1 ;

[0172] Sub-model ZOA-LSTM 0 With sub-model ZOA-LSTM 1 The output, i.e. the prediction result, is fused through the transfer sharing layer, as shown below:

[0173] (15);

[0174] in, To migrate the shared layer output, F For activation functions; For the original sub-model (such as sub-model ZOA-LSTM) 0 The output of ) For the original sub-model (such as sub-model ZOA-LSTM) 0 The weight matrix; For the generation of new sub-models (such as sub-model ZOA-LSTM) 1 The output of ) For the generation of new sub-models (such as sub-model ZOA-LSTM) 1 The weight matrix;

[0175] When acquiring incremental data ΔX 2 At the same time, the sub-model ZOA-LSTM is used in the same way as described above. 1 Perform replication to generate the sub-model ZOA-LSTM 2 Freeze the weights of core layers (such as LSTM hidden layers and feature embedding layers) and use incremental data. ΔX 2 Training sub-model ZOA-LSTM 2 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 2 ;

[0176] Similarly, the sub-model ZOA-LSTM 1 With sub-model ZOA-LSTM 2The output, i.e. the prediction result, is fused through the migration sharing layer; this process is repeated to accumulate and generate a set of sub-models, which utilizes the adaptability of multi-well data to improve the prediction generalization.

[0177] Finally, all sub-models are integrated using error weighting, based on the error weights of each sub-model (ZOA-LSTM). 0 ZOA-LSTM 1 The prediction error of each sub-model (e.g., [model name]) is assigned a weight based on its magnitude in the corresponding well section, and the weighted sum is used to obtain the comprehensive predicted value of net pressure within the fracture in deeper well sections. The core logic is that the smaller the error of the sub-model, the greater its contribution to the final prediction; as shown below:

[0178] (16);

[0179] (17);

[0180] in, Let i be the weight of the i-th sub-model. The error value (such as MAE or other indicators) of the i-th sub-model on the corresponding test set is used to calculate the weights. All weights are normalized and then combined to 1. The predicted value of the i-th sub-model for the target well section (e.g., ZOA-LSTM) 1 (Prediction for 6563~6994 m). This represents the final integrated prediction of the net pressure within the fracture in the target well section, i.e., the output after fusing the prediction results of all sub-models;

[0181] When applying this method, the logging parameters of the target well section are input into each sub-model, and the prediction results are output after weighted fusion; for example... Figure 4-6 As shown, Figure 4 The actual and predicted net pressure values ​​for the shallow well section (4839-6563m) are compared and fitted. Figure 5 A fitting comparison of the actual and predicted net pressure values ​​for the middle well section, i.e., 6563-6994m; Figure 6 The actual and predicted net pressure values ​​for the deep well section (6994-7500m) are compared and fitted.

[0182] Example 3

[0183] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the data-driven method for predicting net pressure within fractured wells as described in Embodiment 1 or 2.

[0184] Example 4

[0185] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the data-driven method for predicting net pressure within fractured wells as described in Embodiment 1 or 2.

[0186] Example 5

[0187] The preprocessing module is configured to: calculate the correlation between logging parameters and net pressure within the fracture and filter them; preprocess the filtered logging parameters and divide them into datasets.

[0188] The model optimization module is configured to: construct an intrasuture net pressure prediction model, input a portion of the dataset into the intrasuture net pressure prediction model for training, and optimize the hyperparameters of the intrasuture net pressure prediction model to obtain an optimized intrasuture net pressure prediction model.

[0189] The model training module is configured to train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model.

[0190] The model prediction module is configured to: when incremental data is acquired, use the trained net pressure prediction model within the fracture as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form a sub-model for different well sections; calculate the comprehensive prediction value of all sub-models to predict the net pressure within the fracture in deeper well sections.

Claims

1. A data-driven method for predicting net pressure within fractured wells, characterized in that, include: Step 1: Collect drilling data, logging parameters, and fracturing operation data from the oilfield site, and calculate the net pressure within the fracture. Step 2: Calculate the correlation between logging parameters and net pressure within the fracture and filter them. Preprocess the filtered logging parameters and divide them into datasets. Step 3: Construct a net intrasuture pressure prediction model. Input a portion of the dataset into the net intrasuture pressure prediction model for training. Optimize the hyperparameters of the net intrasuture pressure prediction model to obtain the optimized net intrasuture pressure prediction model. Step 4: Train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model. Step 5: When acquiring incremental data, use the trained intra-fracture net pressure prediction model as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form sub-models for different well sections; calculate the comprehensive prediction value of all sub-models to predict the intra-fracture net pressure in deeper well sections. The net pressure prediction model within the seam is a long short-term memory neural network, i.e., the LSTM model. A portion of the training set and a portion of the validation set from the dataset are input into the LSTM model for training. During training, the mean absolute percentage error, root mean square error, mean absolute error, and mean square error are normalized and then weighted and summed to form the overall fitness function. The calculation formula is shown below: （14）； in, The comprehensive fitness function is represented by MAPE, the mean absolute percentage error is represented by RMSE, the root mean square error is represented by MAE, and the mean square error is represented by MSE. Optimize the hyperparameters of the LSTM model using the Zebra algorithm; Hyperparameters include the number of hidden layer neurons, learning rate, number of iterations, batch size, and Dropout ratio; Set the zebra population size, where the zebra population represents the set of individuals participating in hyperparameter optimization. Each individual corresponds to a set of hyperparameter combinations. Define the search space boundary, i.e., the range of hyperparameter values. Randomly initialize the zebra population, configure an LSTM model for each zebra individual, and train the LSTM model weights on the training subset (i.e., the training set in a partial dataset). Calculate the Fitness function, i.e., the comprehensive fitness function, as the fitness value on the validation subset (i.e., the validation set in a partial dataset). Select the pioneer zebra with the smallest fitness as the current optimal hyperparameter, and guide the zebra population to update the hyperparameter combination through foraging and defense behaviors; repeat the above iteration until fitness converges or the termination condition is met, output the optimal hyperparameter combination, and obtain the optimized intra-fractal net pressure prediction model. The trained intra-suture net pressure prediction model was used as a sub-model ZOA-LSTM. 0 ; When acquiring incremental data ΔX1, i.e., the logging parameters of the new well section and the corresponding net pressure within the fracture, the ZOA-LSTM sub-model is replicated. 0 Based on the structure and parameters, a new sub-model ZOA-LSTM is generated. 1 Freezing sub-model ZOA-LSTM 1 ZOA-LSTM sub-model 0 Shared core layer weights, only the weight matrix and bias vector parameters of the output layer are allowed to be trained; Train the sub-model ZOA-LSTM using incremental data ΔX1 1 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 1 ; Sub-model ZOA-LSTM 0 With sub-model ZOA-LSTM 1 The output, i.e. the prediction result, is fused through the transfer sharing layer, as shown below: （15）； in, F is the output of the transfer shared layer, and F is the activation function. The output of the original sub-model. This is the weight matrix of the original sub-model; The output of the newly generated sub-model, The weight matrix for the newly generated sub-model; When acquiring incremental data ΔX2, the same operation as above is performed, using the sub-model ZOA-LSTM. 1 Perform replication to generate the sub-model ZOA-LSTM 2 Freeze the core layer weights and train the sub-model ZOA-LSTM using incremental data ΔX2. 2 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 2 ; Similarly, the sub-model ZOA-LSTM 1 With sub-model ZOA-LSTM 2 The output, i.e. the prediction result, is fused through the transfer sharing layer; this process is repeated to accumulate and generate a set of sub-models. Finally, all sub-models are integrated using an error-weighted method. Weights are assigned to each sub-model based on its prediction error in the corresponding well section. The weighted summation yields the comprehensive predicted value of net pressure within the fracture in deeper well sections, as shown below: （16）； （17）； in, Let i be the weight of the i-th sub-model. Let be the error value of the i-th sub-model on the corresponding test set, used to calculate the weights. All weights are normalized and then combined to 1. Let be the predicted value of the i-th sub-model for the target well section. This represents the final integrated prediction of the net pressure within the fracture in the target well section, i.e., the output after fusing the prediction results of all sub-models; When applying the method, the logging parameters of the target well section are input into each sub-model, and the prediction results are output after weighted fusion.

2. The method for predicting net pressure within fractured wells under a data-driven model according to claim 1, characterized in that, Collect on-site drilling data, logging parameters, and fracturing operation data from the oilfield, and calculate the net pressure within the fracture; including: Oilfield drilling data includes vertical depth, tubing inner diameter, and tubing length; Well logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, wellbore diameter, natural gamma, well depth, well inclination angle, and spontaneous potential; Fracturing operation data includes wellhead pressure, hydrostatic column pressure, fracturing fluid friction, near-wellbore friction, and closure stress; The hydrostatic pressure is expressed as: （1）； in, Indicates the hydrostatic pressure; Let g represent the density of the liquid column, g represent the acceleration due to gravity, and h represent the height of the liquid column. Fracturing fluid friction is expressed as: （2）； Where f is the Fanning coefficient of friction. a and b are constants; This refers to the frictional pressure drop caused by viscous friction during the flow of fracturing fluid. Indicates the fracturing fluid flow rate. L is the density of the fracturing fluid; D is the length of the tubing string; p N is the pipeline diameter; Re,gen It is the generalized Reynolds number; Near-wellbore friction is expressed as: （3）； Where, p nw For near-wellbore friction; q is the proppant density; q is the displacement; C d N is the flow coefficient; perf D is the number of holes; perf0 This is the original aperture diameter; The net pressure inside the joint is calculated as follows: （4）； in, This refers to the net pressure within the seam. It is a closed stress; Indicates wellhead pressure, This indicates the hydrostatic pressure.

3. The method for predicting net pressure within fractured wells under a data-driven model according to claim 2, characterized in that, The correlation between well logging parameters and net pressure within the fracture was calculated and filtered. The filtered well logging parameters were preprocessed and divided into datasets, including: The correlation between logging parameters at different depths and the net pressure within the fracture at the corresponding depth was calculated using the Pearson correlation coefficient. The calculation formula is as follows: （5）； in, The Pearson correlation coefficient; Let X be the covariance of Y; and , respectively, are the standard deviations of X and Y; X is the logging parameter, Y is the net pressure within the fracture corresponding to the logging parameter, and n is the sample size, where the sample represents the data samples of logging parameters and corresponding net pressure within the fracture at different depths; , Let X and Y be the observed values ​​in the i-th sample; The average value of X; The average value of Y; Well logging parameters were screened based on the absolute value of the Pearson correlation coefficient, retaining strongly and moderately correlated parameters and eliminating weakly correlated parameters. According to the calculation results, P-wave transit time, S-wave transit time, and rock density were strongly correlated; compensated neutron and wellbore diameter were moderately correlated; and natural gamma, well depth, well inclination angle, and spontaneous potential were weakly correlated. The filtered logging parameters include P-wave transit time, S-wave transit time, rock density, compensated neutrons, and wellbore diameter; Moving average filtering is used to filter and reduce noise in the selected logging parameters, as shown in the following formula: （6）； Among them, X k is the well logging parameter value after moving average filtering; L1 is the width of the moving average window; k is the well logging parameter after the kth filtering; m is the sample size of the moving average window. This represents the mk-th logging parameter; Normalization is performed to normalize the well logging parameter values ​​after moving average filtering to between 0 and 1. The formula is as follows: （7）； in, These are the normalized logging parameter values; These are the well logging parameter values ​​after moving average filtering. and These represent the maximum and minimum values ​​of the logging parameters in the filtered data, respectively. The system is divided into a training set, a validation set, and a test set. The training set consists of normalized logging parameters and net pressure within the fracture. 30% of the data in the training set was randomly selected as the validation set, and the test set consisted of logging parameters and net pressure within the fracture in different well sections.

4. The method for predicting net pressure within fractured wells under a data-driven model according to claim 3, characterized in that, The optimized intra-suture net pressure prediction model was trained using the entire dataset to obtain the trained intra-suture net pressure prediction model. include: All the divided datasets are input into the optimized intra-fracture net pressure prediction model for training. During training, the Adam optimizer is used to minimize the loss function and learn the relationship between the logging parameters in the dataset and the intra-fracture net pressure. The training process was monitored and the intrasuture net pressure prediction model was evaluated using a validation set. The training set was further divided into multiple subsets using cross-validation. After multiple training and validation cycles, the average index was used to evaluate the intrasuture net pressure prediction model. The predicted and actual intrasuture net pressure data were compared, and the performance of the intrasuture net pressure prediction model was evaluated based on any one of the indices: mean absolute percentage error, root mean square error, mean absolute error, and mean square error, as well as the comprehensive fitness function. The training parameters were adjusted until the intrasuture net pressure prediction model reached its optimal performance, thus obtaining the trained intrasuture net pressure prediction model.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the data-driven method for predicting net pressure within fractured wells as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the data-driven method for predicting net pressure within fractured wells as described in any one of claims 1-4.

7. A data-driven system for predicting net pressure within fractured wells, characterized in that, include: The data collection module is configured to collect drilling data, logging parameters, and fracturing operation data from the oilfield, and to calculate the net pressure within the fracture. The preprocessing module is configured to: calculate the correlation between logging parameters and net pressure within the fracture and filter them; preprocess the filtered logging parameters and divide them into datasets. The model optimization module is configured to: construct an intrasuture net pressure prediction model, input a portion of the dataset into the intrasuture net pressure prediction model for training, and optimize the hyperparameters of the intrasuture net pressure prediction model to obtain an optimized intrasuture net pressure prediction model. The model training module is configured to train the optimized intra-suture net pressure prediction model using the entire dataset to obtain the trained intra-suture net pressure prediction model. The model prediction module is configured to: when incremental data is acquired, use the trained net pressure prediction model within the fracture as a sub-model, copy the structure and parameters of the sub-model to generate a new sub-model, and train it to form a sub-model for different well sections; calculate the comprehensive prediction value of all sub-models to predict the net pressure within the fracture in deeper well sections; The net pressure prediction model within the seam is a long short-term memory neural network, i.e., the LSTM model. A portion of the training set and a portion of the validation set from the dataset are input into the LSTM model for training. During training, the mean absolute percentage error, root mean square error, mean absolute error, and mean square error are normalized and then weighted and summed to form the overall fitness function. The calculation formula is shown below: （14）； in, The comprehensive fitness function is represented by MAPE, the mean absolute percentage error is represented by RMSE, the root mean square error is represented by MAE, and the mean square error is represented by MSE. Optimize the hyperparameters of the LSTM model using the Zebra algorithm; Hyperparameters include the number of hidden layer neurons, learning rate, number of iterations, batch size, and Dropout ratio; Set the zebra population size, where the zebra population represents the set of individuals participating in hyperparameter optimization. Each individual corresponds to a set of hyperparameter combinations. Define the search space boundary, i.e., the range of hyperparameter values. Randomly initialize the zebra population, configure an LSTM model for each zebra individual, and train the LSTM model weights on the training subset (i.e., the training set in a partial dataset). Calculate the Fitness function, i.e., the comprehensive fitness function, as the fitness value on the validation subset (i.e., the validation set in a partial dataset). Select the pioneer zebra with the smallest fitness as the current optimal hyperparameter, and guide the zebra population to update the hyperparameter combination through foraging and defense behaviors; repeat the above iteration until fitness converges or the termination condition is met, output the optimal hyperparameter combination, and obtain the optimized intra-fractal net pressure prediction model. The trained intra-suture net pressure prediction model was used as a sub-model ZOA-LSTM. 0 ; When acquiring incremental data ΔX1, i.e., the logging parameters of the new well section and the corresponding net pressure within the fracture, the ZOA-LSTM sub-model is replicated. 0 Based on the structure and parameters, a new sub-model ZOA-LSTM is generated. 1 Freezing sub-model ZOA-LSTM 1 ZOA-LSTM sub-model 0 Shared core layer weights, only the weight matrix and bias vector parameters of the output layer are allowed to be trained; Train the sub-model ZOA-LSTM using incremental data ΔX1 1 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 1 ; Sub-model ZOA-LSTM 0 With sub-model ZOA-LSTM 1 The output, i.e. the prediction result, is fused through the transfer sharing layer, as shown below: （15）； in, F is the output of the transfer shared layer, and F is the activation function. The output of the original sub-model. This is the weight matrix of the original sub-model; The output of the newly generated sub-model, The weight matrix for the newly generated sub-model; When acquiring incremental data ΔX2, the same operation as above is performed, using the sub-model ZOA-LSTM. 1 Perform replication to generate the sub-model ZOA-LSTM 2 Freeze the core layer weights and train the sub-model ZOA-LSTM using incremental data ΔX2. 2 The trainable layers are used to obtain the sub-model ZOA-LSTM adapted to the new well section. 2 ; Similarly, the sub-model ZOA-LSTM 1 With sub-model ZOA-LSTM 2 The output, i.e. the prediction result, is fused through the transfer sharing layer; this process is repeated to accumulate and generate a set of sub-models. Finally, all sub-models are integrated using an error-weighted method. Weights are assigned to each sub-model based on its prediction error in the corresponding well section. The weighted summation yields the comprehensive predicted value of net pressure within the fracture in deeper well sections, as shown below: （16）； （17）； in, Let i be the weight of the i-th sub-model. Let be the error value of the i-th sub-model on the corresponding test set, used to calculate the weights. All weights are normalized and then combined to 1. Let be the predicted value of the i-th sub-model for the target well section. This represents the final integrated prediction of the net pressure within the fracture in the target well section, i.e., the output after fusing the prediction results of all sub-models; When applying the method, the logging parameters of the target well section are input into each sub-model, and the prediction results are output after weighted fusion.

Images