A Machine Learning Model-Based Method for Predicting Fracture Pressure in Layered Shale Wells with Large Distance Distance

Abstract

This invention relates to a method for predicting fracture pressure in layered shale wells with extended reach based on a machine learning model. S1: Acquire drilled well data and collect downhole core samples, conduct anisotropic tensile mechanical experiments on shale along the bedding direction, and construct shale fracture criteria based on the experimental results; S2: Construct a fracture pressure model for layered shale wells with extended reach, and correct and integrate the calculated fracture pressure values ​​with field fracture test data; S3: Calculate the fracture pressure using a mechanical model based on the wellbore structure conditions of highly deviated wells, verify it using field low-pressure data, and form a fracture pressure calculation and analysis database; S4: Train the prediction model based on machine learning methods, determine the WT-LSTM model structure, set hidden layers, and train it; S5: Optimize the machine learning-based fracture pressure model through hyperparameter adjustment, feature selection, and importance assessment. This invention comprehensively improves the accuracy of fracture pressure prediction in highly deviated wells by integrating multiple factors, provides data support for the design of fracturing engineering in complex layered shale, helps to rationally select parameters for extended reach well construction, improves adaptability and safety, and provides data support for the development of layered shale oil and gas resources.

Description

Technical Field

[0001] This invention belongs to the field of oilfield development technology, specifically relating to a method for predicting fracture pressure in layered shale wells with large displacement based on machine learning models. Background Technology

[0002] With my country's ever-increasing energy demand, the development of conventional oil and gas fields has entered its later stages, and the development of unconventional oil and gas resources such as shale is receiving increasing attention. my country is rich in shale oil and gas resources, and horizontal well hydraulic fracturing technology is an effective measure for developing shale oil and gas. However, my country's geological structure is complex, and shale reservoirs are characterized by large thickness, small single-layer thickness, significant sweet spot thin interbedded features, and numerous stress-isolated interlayers. Traditional horizontal well volumetric fracturing technology has limited vertical penetration effect. Therefore, the use of highly deviated wells for fracturing development of shale reservoirs has been proposed.

[0003] Fracturing pressure is a crucial parameter in fracturing and drilling engineering design. Fracturing in highly deviated wells differs from that in vertical and horizontal wells. Under given stress field conditions, highly deviated wells are affected by factors such as well inclination angle, wellbore azimuth angle, and perforation phase angle. After fracture initiation, the fractures first twist and deflect before extending along the direction of maximum principal stress, resulting in complex fracture propagation morphology. Furthermore, after initiation, the fractures primarily form single-fracture serrated fractures or multiple fractures perpendicular to the wellbore. Fracture twisting increases the friction of the fracturing fluid, making proppant addition difficult. Therefore, accurately calculating the fracturing pressure of shale has always been a challenging problem in oil and gas engineering.

[0004] Shale strata are well-developed, and therefore their mechanical properties tend to exhibit different mechanical characteristics along different bedding directions. In the past, many scholars have conducted extensive research on the tensile mechanical properties of bedding rocks.

[0005] Existing fracture pressure models mainly include elasticity models and fracture mechanics models. Meanwhile, selecting a suitable rock fracture criterion is also crucial to the fracture pressure calculation results. Due to the influence of well inclination and azimuth angles, it is difficult to directly calculate using well logging data; typically, modeling and obtaining fracture pressure values ​​at a fixed depth is required, a relatively cumbersome process. Machine learning technology is a data-driven prediction method with stronger generalization ability compared to theoretical models and has been applied in the oil and gas exploration and development industry. Based on the differences in the required tasks, common machine learning methods can be divided into supervised learning, unsupervised learning, etc. Supervised learning is mainly represented by artificial neural networks, while unsupervised learning mainly focuses on solving clustering and dimensionality reduction problems. Other machine learning methods include time series analysis, whose advantage is that it can apply relevant historical data to recurrent neural networks to better utilize historical information. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting fracture pressure in layered shale wells with large displacement based on machine learning models.

[0007] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for predicting fracture pressure in layered shale wells with large displacement based on machine learning models, the steps of which are as follows:

[0008] S1. Obtain logging and core data from drilled wells. Core samples are taken along different bedding angles. After removing residual air from the cores and curing them, anisotropic tensile strength tests are conducted on the bedding shale to obtain the characteristics of tensile strength variation under different bedding angles, and then an anisotropic fracture criterion for shale is constructed.

[0009] S2. Based on the tensile strength of rock under different bedding angles obtained in S1, the transverse isotropic fracture pressure criterion is selected and the fracture pressure is calculated. At the same time, an anisotropic geostress field combined spring model of bedding shale is constructed, and the calculated fracture pressure value is corrected and integrated with the field ground failure test data to complete the fracture pressure modeling of bedding shale large displacement well;

[0010] S3. Record the fracture pressure variation under different wellbore structural parameters (inclination angle, azimuth angle) and formation parameters (vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, pore pressure). Compare and verify the numerical simulation results with relevant data from hydraulically fractured wells already implemented in the oilfield block. Evaluate the accuracy of the calculation model by calculating the average error index, and make necessary adjustments based on the evaluation results. Then, perform standardization processing, and divide the processed data into training and test sets according to a certain ratio to form a fracture pressure database for layered shale high-displacement wells.

[0011] S4. Determine the WT-LSTM model structure, set the number of hidden layers and nodes in each layer of the LSTM, initialize the weights and bias parameters of the LSTM part using normal or uniform distribution, select mean squared error to measure the difference between the predicted and true values, update the model parameters using mini-batch stochastic gradient descent algorithm, use a learning rate decay strategy, combine L2 regularization and Dropout techniques to prevent overfitting, adjust the training process based on the validation set index, and obtain the machine learning prediction model.

[0012] S5. The hyperparameters are finely tuned by combining grid search or random search with cross-validation, and the optimal hyperparameter combination is selected to update the model. The feature importance evaluation method in random forest is used to re-evaluate the importance of input features, remove features with small contributions and try new feature combinations. Multiple different neural network models are trained, and the prediction results are integrated by weighted averaging to improve accuracy and stability. Finally, the optimized machine learning model is used to predict the fracture pressure of large-displacement wells in rational shale formations.

[0013] Furthermore, S1 also includes:

[0014] By taking core samples along different bedding angles, then vacuuming the cores to remove residual air and curing them, tensile strength tests were conducted on the bedding shale to obtain the characteristics of tensile strength variation of the bedding shale under different bedding angles, providing data support for the subsequent construction of bedding models.

[0015] Furthermore, S2 also includes:

[0016] S21. Based on the tensile strength of the bedding shale obtained in S1 under different bedding angles, the transverse isotropic fracture pressure criterion is optimized, where the calculation formula for the single weak surface criterion is as follows:

[0017]

[0018] In the formula: S t (θ) represents the tensile strength under different bedding angles, in MPa; θ is the angle between the bedding plane and the horizontal stress, in °; T b T represents the tensile strength of the bedding plane, in MPa; m θ represents the tensile strength of the rock matrix, in MPa; * The critical angle is °, derived from... Sure.

[0019] Based on the above criteria, the rupture pressure can be calculated.

[0020] P f (θ)=3σ h -σ H -αp p +S t (θ) (Equation 2)

[0021] In the formula: P f (θ) represents the fracture pressure under different bedding angles, in MPa; σ H and σ h These represent the maximum and minimum horizontal ground stresses, respectively, in MPa; α is the Biot coefficient; P p Formation pore pressure; S t (θ) represents the tensile strength under different bedding angles, in MPa;

[0022] The calculated rupture pressure value usually needs to be corrected and integrated with field ground rupture test data to ensure that the rupture pressure value calculated by the mechanism model conforms to the actual field conditions.

[0023] S22. An anisotropic geostress field combined spring model needs to be constructed, as shown below:

[0024] 1. Elastic parameter relationships: For strata containing a single weak plane (assuming the plane lies in the XOY plane), the Poisson's ratio and elastic modulus of anisotropic elastic bodies exhibit specific relationships in different directions, such as:

[0025]

[0026] Where: μ H and E H μ represents Poisson's ratio and elastic modulus of the parallel plane, respectively. ν and E ν These are Poisson's ratio and elastic modulus at the vertical level, respectively.

[0027] 2. Generalized Hooke's Law Expression: The generalized Hooke's Law describes the relationship between principal stresses and strain in a transversely isotropic elastic body. Its expression is as follows:

[0028]

[0029] In this formula, Where, σ H σ h and σ v These represent the maximum horizontal stress, minimum horizontal stress, and vertical stress, respectively; α is the Biot coefficient; P p This refers to pore pressure.

[0030] 3. Calculation of In-situ Stress Using a Combined Spring Model: During tectonic movements, the deformation effects of each tectonic movement on the strata are coordinated, meaning the deformation amount of each layer is the same. However, due to differences in stiffness (elastic modulus and Poisson's ratio) among the layers, the internal tectonic stresses generated are different. Based on this principle, the σo in transversely isotropic strata... H and σ h It can be calculated using the following formula:

[0031]

[0032] Furthermore, S3 also includes:

[0033] S31. Based on the wellbore structure conditions of layered shale wells with extended reach, and combined with the previously established anisotropic fracture criterion and related physical models, the fracture pressure calculation for layered shale wells with extended reach is carried out. Numerical simulation methods are used to simulate and calculate the wellbore stress distribution under different wellbore structure conditions, thereby determining the fracture pressure. During the calculation process, the variation law of fracture pressure under different wellbore structure parameters (inclination angle, azimuth angle) and formation parameters (vertical depth, elastic modulus, Poisson's ratio, in-situ stress, porosity, tensile strength, pore pressure) is recorded. The obtained fracture pressure values ​​are compared and verified with the actual collected data.

[0034] S32. The accuracy of the computational model can be evaluated using the mean error index (ME). (If ME is close to 0, it indicates that the average deviation of the model's predicted values ​​is small, and the model's accuracy is high. When ME is positive, it means that the model's overall predicted values ​​are too high; when it is negative, it means that the model's overall predicted values ​​are too low. Generally speaking, the smaller the absolute value of ME, the better. For example, in some simple numerical prediction scenarios, if the absolute value of ME is less than 5%-10% of the true value range, the model can be considered to have a certain degree of accuracy, but this is not an absolute standard and needs to be analyzed in conjunction with the specific problem.) If the error is large, the computational model and input parameters need to be checked, and necessary corrections and adjustments need to be made.

[0035] The formula for calculating the average error is:

[0036]

[0037] In the formula: n represents the number of samples, that is, the total number of data points used to calculate the average error; y is the model's prediction for the i-th sample; i Let be the true value of the i-th sample.

[0038] S33. Standardize the wellbore structural parameters (inclination angle, azimuth angle), formation parameters (vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, pore pressure), and fracture pressure values. Normalize or standardize input features with different dimensions and numerical ranges. Normalization methods can be used to map data to the 0-1 interval, improving the convergence speed and accuracy of machine learning algorithms and preventing certain features from having a dominant influence on model training due to excessively large or small values.

[0039] S34. Divide the processed data into training and testing sets according to a certain ratio (e.g., 70%-80% for training and 20%-30% for testing) to generate a dataset for subsequent model training and evaluation.

[0040] Furthermore, S4 also includes:

[0041] S41. Determine the WT-LSTM model structure. First, the number of nodes in the input layer is determined based on the number of selected input features. Nine features are selected: well inclination angle, azimuth angle, vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, and pore pressure. The number of nodes in the input layer is 9. In the WT-LSTM model, the input data is first decomposed into subsequences of different frequency components through wavelet transform. These subsequences are then input into a Long Short-Term Memory (LSTM) network. The number of nodes in the output layer remains at 1, used to output the predicted fracture pressure value of the layered shale.

[0042] S42. The hidden layer needs to be configured as follows:

[0043] 1. Determining the Number of LSTM Hidden Layers: Given the characteristics of LSTM in the WT-LSTM model, we started with a single LSTM hidden layer. We constructed the model using a preliminary estimate of the number of nodes based on the number of input features and problem complexity. We trained the model on the training set and evaluated its performance on the validation set (using root mean square error and accuracy as indicators). If the performance did not meet expectations, we gradually increased the number of LSTM hidden layers, repeating the training and evaluation process after each increase, while monitoring the model's performance trend on the validation set. We applied L2 regularization (setting an initial regularization coefficient) to control model complexity and prevent overfitting when increasing the number of LSTM hidden layers. Based on performance and overfitting, we determined a suitable range for the number of LSTM hidden layers.

[0044] 2. Determining the Number of LSTM Hidden Layer Nodes (for a predetermined number of LSTM hidden layers): For each LSTM hidden layer, the initial range of the number of nodes is determined based on the number of input and output features. In this patent, when there are 9 input features, the initial range of the number of nodes in the first LSTM hidden layer is set to 10-30 (step size 5). If a second LSTM hidden layer is determined, the initial range of its number of nodes is set to 5-20 (step size 5). Using grid search or random search methods, various combinations are traversed within the set number of nodes. After building the model, it is trained on the training set, and its performance is evaluated on the validation set. The performance changes of the model under different combinations of node numbers are observed, and the combination of node numbers with the best performance on the validation set is selected. If the performance difference is small (RMSE difference within 0.1-0.5 MPa is considered insignificant), the combination with fewer nodes is preferred to reduce computational cost and the risk of overfitting.

[0045] S43. The specific steps for model training are as follows:

[0046] 1. For the weights and bias parameters of the LSTM part in the WT-LSTM model, use a normal or uniform distribution for initialization and set them to small values ​​(close to 0) to avoid instability caused by excessive gradients in the early stage of training.

[0047] 2. The mean squared error (MSE) is used to measure the difference between the model's predicted value and the true value. The model parameters are optimized by minimizing the loss function, so that the model's prediction results are as close as possible to the true value.

[0048] 3. Update the LSTM model parameters using the mini-batch stochastic gradient descent (SGD) algorithm. Divide the wavelet transform preprocessed data into several mini-batches (e.g., each mini-batch contains 32 samples), and calculate the gradient of the LSTM part using one mini-batch of data at a time. Calculate the gradient of the loss function with respect to the LSTM model parameters, and then update the parameters according to the formula.

[0049]

[0050] In the formula: θ is the model parameter, α is the learning rate, m is the number of mini-batch samples, and ▽J(θ) old ,x i ,y i ) is the gradient of the loss function for the i-th sample within a mini-batch with respect to the model parameters.

[0051] 4. During training, use a learning rate decay strategy. For example, use exponential decay, setting the initial learning rate α0 = 0.01 and the decay factor β = 0.95. At each iteration t, according to the formula α... t =α0β t Calculate the current learning rate and gradually decrease it as training progresses, allowing the model to finely adjust its parameters.

[0052] 5. To prevent overfitting, L2 regularization is used. A regularization term is added to the loss function. (Where λ is the regularization coefficient and W is the set of model weights), the model complexity is limited by penalizing the magnitude of the model weights. Simultaneously, Dropout technology can be combined to randomly drop some hidden layer neurons during training with a certain probability (e.g., 0.5), increasing the model's robustness.

[0053] 6. After each epoch of training (one iteration of the training set), calculate the root mean square error (RMSE) and coefficient of determination (R²) of the model on the validation set. 2 Evaluation metrics such as RMSE, RMSE, etc. should be closely monitored. If the RMSE of the validation set no longer shows a decreasing or increasing trend for several consecutive epochs (5 epochs), training should be terminated early to prevent overfitting and select the model with the best performance.

[0054] 7. Repeat steps 3-6 until the training stopping condition is met (e.g., reaching the preset maximum number of training iterations, such as 200 epochs, or the validation set performance no longer improves). Save the best model obtained from training, including the model structure and parameters, for subsequent prediction applications.

[0055] Furthermore, S5 also includes:

[0056] S51. Hyperparameter Tuning: Fine-tuning the hyperparameters of the trained model, such as the learning rate, regularization coefficient, and number of hidden layer nodes. A grid search or random search combined with cross-validation is used to search within a wider range of parameter values ​​to find the optimal hyperparameter combination for model performance. Based on the previously determined initial range for the number of hidden layer nodes, the search step size is further refined. For the learning rate, a smaller step size (e.g., 0.001) is used within the range [0.001, 0.1], and for the regularization coefficient, the search is performed within the range [0.01, 0.1]. Through multiple cross-validations (5-fold cross-validation), the performance of the model on the validation set under different hyperparameter combinations (e.g., root mean square error, coefficient of determination) is evaluated, and the best-performing hyperparameter combination is selected to update the model.

[0057] S52. Feature Selection and Importance Assessment: Re-evaluate the importance of input features using a feature importance assessment algorithm (feature importance assessment method in random forests). Based on the feature importance scores, remove features that contribute little to the model's predictions, retaining key features to reduce model complexity and improve computational efficiency. If a geological parameter feature has an extremely low importance score for rupture pressure prediction, consider removing it from the input features and then retrain the model to observe performance changes. Simultaneously, try combining or transforming some features, such as multiplying or adding certain related features to obtain new features, and re-evaluate the impact of the new feature combinations on model performance, selecting feature combination methods that improve model performance.

[0058] S53. Model Ensemble: Model ensemble techniques are used to improve prediction accuracy and stability. Multiple neural network models with different initial parameters or structures (e.g., slightly different numbers of hidden layers) are trained, and then the prediction results of these models are integrated using a weighted average method. In the weighted average method, the weights of each model are determined based on its performance on the validation set, with higher-performing models assigned larger weights. Model ensemble effectively reduces the error of individual models and improves the overall reliability of predictions.

[0059] The beneficial effects of this invention are as follows: by acquiring multi-source data to conduct experiments and constructing anisotropic fracture criteria, and combining this with wellbore structural conditions to calculate and verify a database, the accuracy of fracture pressure prediction is improved by comprehensively considering multiple factors, providing precise data support for engineering design; the constructed relevant physical models make the calculations more scientific and reasonable, and recording and verifying patterns during the calculations helps ensure accuracy; the screening and processing of data to generate high-quality datasets and the optimization of model structure and training process improve calculation accuracy and efficiency, enabling the model to better learn features for accurate prediction; accurate prediction helps to rationally select parameters and control pressure during the construction of extended reach wells, avoid construction problems, and improve adaptability and safety; ultimately, it provides a basis for engineering decisions and assists in resource development. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 A flowchart illustrating a method for predicting fracture pressure in layered shale wells with extended reach based on a machine learning model.

[0062] Figure 2 Diagram showing the relationship between circumferential stress and wellbore angle under different bedding angles and tensile strengths.

[0063] Figure 3 A schematic diagram illustrating the calculation of fracture pressure under well inclination angle of 45° and phase angle of 0°.

[0064] Figure 4 A schematic diagram illustrating the calculation of fracture pressure under well inclination angle and phase angle conditions of 45°.

[0065] Figure 5 A schematic diagram illustrating the calculation of fracture pressure under well inclination angles of 45° and phase angles of 90°.

[0066] Figure 6 A schematic diagram illustrating the variation in calculation error of formation fracture pressure under different hidden layer conditions.

[0067] Figure 7 A comparison table of in-situ ground failure experimental data, rupture pressure data from mechanism calculation models, and rupture pressure from machine learning-driven calculation models under different depth conditions.

Claims

1. A method for predicting fracture pressure in layered shale wells with large displacement based on a machine learning model, characterized in that, include: S1. Obtain logging and core data from drilled wells; take core samples along different bedding angles, remove residual air from the core samples by vacuuming and curing them, and then conduct anisotropic tensile strength tests on bedding shale to obtain the characteristics of rock tensile strength variation under different bedding angles, and then construct anisotropic fracture criteria for shale. S2. The tensile strength of the layered shale under different bedding angles obtained from S1 is used to select the transverse isotropic fracture pressure criterion; at the same time, an anisotropic geostress field combined spring model of layered shale is constructed, and the calculated fracture pressure value is corrected and integrated by combining the field ground failure test data to complete the fracture pressure modeling of layered shale large displacement well. S3. Record wellbore structural parameters, including inclination angle and azimuth angle, as well as formation parameters, including vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, and pore pressure, and the variation law of fracture pressure under these conditions; compare and verify the numerical simulation calculations with relevant data from hydraulically fractured wells already implemented in the oilfield block; evaluate the accuracy of the calculation model by calculating the average error index; and make corrections and adjustments based on the evaluation results; then perform standardization processing, and divide the processed data into training and test sets according to a certain proportion to form a fracture pressure database for layered shale high-displacement wells; S4. Determine the WT-LSTM model structure, set the number of LSTM hidden layers and the number of nodes in each layer, initialize the weights and bias parameters of the LSTM part using normal or uniform distribution, select mean squared error to measure the difference between the predicted value and the true value, use mini-batch stochastic gradient descent algorithm to update the model parameters, use the learning rate decay strategy, combine L2 regularization and Dropout technology to prevent overfitting, adjust the training process according to the validation set index, and obtain the machine learning prediction model. S5. The hyperparameters are finely tuned by combining grid search or random search with cross-validation, and the optimal hyperparameter combination is selected to update the model. The feature importance evaluation method in random forest is used to re-evaluate the importance of input features, remove features with small contributions and try new feature combinations. Multiple different neural network models are trained, and the prediction results are integrated by weighted averaging to improve accuracy and stability. Finally, the optimized machine learning model is used to predict the fracture pressure of large-displacement wells in rational shale formations.

2. The method for predicting fracture pressure in layered shale wells with large displacement based on a machine learning model according to claim 1, wherein step S1 further includes: By taking core samples along different bedding angles, then vacuuming the cores to remove residual air and curing them, tensile strength tests were conducted on the bedding shale to obtain the characteristics of tensile strength variation of the bedding shale under different bedding angles, providing data support for the subsequent construction of bedding models.

3. The method for predicting fracture pressure in layered shale wells with large displacement based on a machine learning model according to claim 1, wherein step S3 further includes: S31. Based on the wellbore structure conditions of layered shale wells with large reachability, and combined with the previously constructed anisotropic fracture criterion and related physical models, the fracture pressure of layered shale wells with large reachability is calculated. Numerical simulation methods are used to simulate and calculate the wellbore stress distribution under different wellbore structure conditions, thereby determining the fracture pressure. During the calculation process, wellbore structure parameters, including inclination angle and azimuth angle, as well as formation parameters, including vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, and pore pressure, are recorded, and the fracture pressure variation law is observed. The obtained fracture pressure values ​​are compared and verified with the actual collected data. S32. The mean error index (ME) is used to evaluate the accuracy of the computational model. When ME is positive, it indicates that the model's overall predictions are too high; when ME is negative, it indicates that the model's overall predictions are too low. The smaller the absolute value of ME, the higher the model's accuracy. When the ME error is large, the computational model and input parameters need to be checked and corrected or adjusted. The formula for calculating the average error is: In the formula: n represents the number of samples, that is, the total number of data points used to calculate the average error; y is the model's prediction for the i-th sample; i This represents the true value of the i-th sample. S33. Standardize the wellbore structural parameters, including inclination angle and azimuth angle, and the formation parameters, including vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, and pore pressure, as well as the fracture pressure value. Normalize or standardize the input features with different dimensions and numerical ranges. Use the normalization method to map the data to the 0-1 interval to improve the convergence speed and accuracy of the machine learning algorithm and avoid certain features from having a dominant influence on model training due to excessively large or small values. S34. Divide the processed data into training and testing sets according to the ratio of 70%-80% as training set and 20%-30% as testing set, and generate a dataset for subsequent model training and evaluation.

4. The method for predicting fracture pressure in layered shale wells with large displacement based on a machine learning model according to claim 1, wherein step S4 further includes: S41. Determine the WT-LSTM model structure: The number of input layer nodes is determined based on the number of selected input features. Nine features are selected: well inclination angle, azimuth angle, vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, and pore pressure. The number of input layer nodes is 9. In the WT-LSTM model, the input data is first decomposed into subsequences of different frequency components by wavelet transform. These subsequences are then input into the Long Short-Term Memory (LSTM) network. The number of output layer nodes remains at 1, used to output the predicted fracture pressure value of the layered shale. S42. The hidden layer needs to be configured as follows: Determining the Number of LSTM Hidden Layers: Given the characteristics of LSTM in the WT-LSTM model, we started with one LSTM hidden layer. We constructed the model using a preliminary estimate of the number of nodes based on the number of input features and problem complexity. We trained the model on the training set and evaluated its performance on the validation set using root mean square error and accuracy as metrics. If the performance did not meet expectations, we gradually increased the number of LSTM hidden layers, repeating the training and evaluation process after each increase, while monitoring the model's performance trend on the validation set. We applied L2 regularization and set an initial regularization coefficient to control model complexity and prevent overfitting when increasing the number of LSTM hidden layers. Based on performance and overfitting, we determined a suitable range for the number of LSTM hidden layers. S43. The specific steps for model training are as follows: a. Initialize the weights and bias parameters of the LSTM part in the WT-LSTM model using a normal or uniform distribution; b. Use mean squared error (MSE) to measure the difference between the model's predicted value and the true value, and use the minimization of the loss function to optimize the model parameters, so that the model's prediction results are closer to the true value. c. Update the model parameters of the LSTM part using the mini-batch stochastic gradient descent (SGD) algorithm; calculate the gradient of the LSTM part using a mini-batch of data each time; calculate the gradient of the loss function with respect to the LSTM model parameters, and then update the parameters according to the formula. In the formula: For model parameters, Here, m is the learning rate, and m is the number of samples in the mini-batch. Let be the gradient of the loss function for the i-th sample within a mini-batch with respect to the model parameters; d. During training, a learning rate decay strategy is used; that is, an exponential decay method is adopted to set the initial learning rate. The attenuation factor β = 0.

95. At each iteration t, according to the formula... Calculate the current learning rate and gradually reduce it as training progresses, allowing the model to finely adjust its parameters. e. To prevent overfitting, L2 regularization is used; a regularization term is added to the loss function. Where λ is the regularization coefficient and W is the set of model weights, the model complexity is limited by penalizing the size of the model weights; at the same time, combined with the Dropout technique, some hidden layer neurons are randomly dropped with a probability of 0.5 during training to increase the robustness of the model. f. After completing each epoch (i.e., traversing the training set once), calculate the root mean square error (RMSE) and coefficient of determination (R²) of the model on the validation set. 2 Evaluation metrics; closely monitor changes in these metrics. If the validation set RMSE no longer shows a decreasing or increasing trend for 5 consecutive epochs, terminate training early to prevent overfitting and select the best-performing model. Repeat steps c to f until the training stopping condition is met, reaching the preset maximum number of training epochs of 200, or until the performance on the validation set no longer improves. Save the best model obtained from training, including the model structure and parameters, for subsequent prediction applications.

5. The method for predicting fracture pressure in layered shale wells with large displacement based on a machine learning model according to claim 1, wherein step S5 further includes: S51. Hyperparameter Tuning: Fine-tuning the hyperparameters of the trained model, including the learning rate, regularization coefficient, and number of hidden layer nodes. A grid search or random search combined with cross-validation method is used to search within a wider range of parameter values ​​to find the hyperparameter combination that optimizes model performance. Based on the previously determined initial range of hidden layer node numbers, the search step size is further refined: the learning rate is searched with a step size of 0.001 in the range of 0.001 to 0.1, and the regularization coefficient is searched in the range of 0.01 to 0.

1. Through multiple 5-fold cross-validations, the root mean square error and coefficient of determination are used to evaluate the model's performance on the validation set under different hyperparameter combinations, and the hyperparameter combination with the best performance is selected to update the model. S52, Feature Selection and Importance Assessment: Re-evaluate the importance of input features by using feature importance assessment methods from random forests; Based on feature importance scores, features that contribute little to the model's prediction are removed, while key features are retained to reduce model complexity and improve computational efficiency. If the importance score of a certain geological parameter feature for rupture pressure prediction is lower than a set threshold, it is removed from the input features, and the model is retrained to observe performance changes. At the same time, some features are combined or transformed, including multiplying or adding certain related features to obtain new features. The impact of the new feature combination on model performance is evaluated again, and feature combination methods that can improve model performance are selected. S53. Model Ensemble: Model ensemble techniques are used to improve prediction accuracy and stability. Multiple neural network models with different initial parameters or different structures are trained, and then the prediction results of these models are integrated using a weighted average method. In the weighted average method, the weights of each model are determined based on its performance on the validation set, and models with better performance are given larger weights. Through model ensemble, the error of a single model can be effectively reduced, and the reliability of the overall prediction can be improved.

Images