Method for predicting fracture pressure of stratified shale extended reach well based on machine learning model

Abstract

The invention relates to a method for predicting fracture pressure of a stratified shale extended reach well based on a machine learning model. The method comprises the following steps: S1, acquiring well drilling data, collecting an underground rock core, carrying out an anisotropic shale tensile mechanical experiment along a bedding direction, and constructing a shale fracture criterion according to an experiment result; s2, constructing a stratified shale extended reach well fracture pressure model, and correcting and integrating the calculated fracture pressure value in combination with field ground fracture experimental data; s3, combining well body structure conditions of the highly-deviated well, using a mechanical model to calculate fracture pressure, and using field small pressure data verification to form a fracture pressure calculation and analysis database; s4, training the prediction model based on a machine learning method, determining a WT-LSTM model structure, setting a hidden layer and performing training; and S5, optimizing the rupture pressure model based on machine learning through hyper-parameter adjustment, feature selection and importance evaluation means. According to the method, the prediction accuracy of the fracture pressure of the highly-deviated well is improved by integrating multiple factors, data support is provided for on-site complex stratified shale fracturing engineering design, reasonable parameter selection of extended reach well construction is facilitated, adaptability and safety are improved, and data support is provided for stratified shale oil and gas resource development.

Description

Technical Field

[0001] The present invention belongs to the technical field of oilfield development, and in particular relates to a method for predicting the fracture pressure of a stratified shale extended-reach well based on a machine learning model. Background Art

[0002] As my country's demand for energy continues to increase, the development of conventional oil and gas fields has entered the late stage, and the development of unconventional oil and gas resources such as shale has received increasing attention. my country is rich in shale oil and gas resources, and horizontal well hydraulic fracturing technology is an effective measure to develop 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 interlayer characteristics of sweet spots, and many stress isolation interlayers. The vertical penetration effect of traditional horizontal well volume fracturing technology is limited. Therefore, it is proposed to use high-angle wells for fracturing development of shale reservoirs.

[0003] The fracture pressure is an important parameter for fracturing and drilling engineering design. The fracturing of highly deviated wells is different from that of vertical wells and horizontal wells. Under given stress field conditions, highly deviated wells are affected by the well inclination angle, wellbore azimuth angle, and perforation phase angle. After the fracture is initiated, it first twists and deflects, and then extends along the direction of the maximum principal stress, resulting in a complex fracture expansion morphology. After the fracture is initiated, it mainly forms a single-fracture-dominated serrated fracture or multiple fractures perpendicular to the wellbore. The distortion of the fracture increases the friction of the fracturing fluid and makes it difficult to add sand. Therefore, the accurate calculation of shale fracture pressure has always been a difficult problem in oil and gas engineering.

[0004] Shale has well-developed stratification, so its mechanical properties tend to show different mechanical characteristics in different bedding directions. In the past, many scholars have conducted extensive research on the tensile mechanical properties of stratified rocks.

[0005] The existing fracture pressure models mainly include: elastic mechanics model and fracture mechanics model. At present. At the same time, the selection of appropriate rock fracture criteria is also the key to affecting the fracture pressure calculation results. Due to the influence of well inclination and azimuth, it is difficult to directly use logging data for calculation. It is usually necessary to model and obtain the fracture pressure value at a fixed depth, and the process is relatively cumbersome. Machine learning technology is a data-driven prediction method. Compared with theoretical models, it has stronger generalization ability and has been applied to the oil and gas exploration and development industry. According to the differences in the required tasks, common machine learning methods can be divided into supervised learning, unsupervised learning and other categories. Among them, supervised learning is mainly represented by artificial neural networks, and unsupervised learning mainly solves clustering and dimensionality reduction problems. In addition, there are machine learning methods such as time series analysis. Its advantage is that relevant data in history can be applied to recurrent neural networks to better use the historical information of the data. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a method for predicting the fracture pressure of a stratified shale extended-reach well based on a machine learning model.

[0007] The technical solution provided by the present invention to solve the above technical problems is: a method for predicting the fracture pressure of a stratified shale extended reach well based on a machine learning model, the steps of which are as follows:

[0008] S1. Obtain the logging and core data of the drilled wells. Take cores along different bedding angles, vacuum the cores to remove residual air and maintain them, then conduct anisotropic tensile strength tests on bedding shale to obtain the variation characteristics of rock tensile strength under different bedding angles, and then construct anisotropic fracture criteria for shale;

[0009] S2. Based on the rock tensile strength 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, a combined spring model of the anisotropic geostress field of the layered shale is constructed, and the calculated fracture pressure value is corrected and integrated in combination with the field ground failure experimental data to complete the fracture pressure modeling of the layered shale large displacement well;

[0010] S3. Record the variation patterns of fracture pressure under different wellbore structural parameters (well 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 calculation values ​​with the relevant data of hydraulic fracturing wells in the oil field block, evaluate the accuracy of the calculation model by calculating the average error index, and make corrections and adjustments when necessary based on the evaluation results. Then perform standardization processing, divide the processed data into training sets and test sets according to a certain ratio, and form a fracture pressure database for stratigraphic shale large-displacement wells;

[0011] S4. Determine the WT-LSTM model structure, set the number of LSTM hidden layers and the number of nodes in each layer, use normal or uniform distribution to initialize the weights and bias parameters of the LSTM part, use mean square error to measure the difference between the predicted value and the true value, use the small 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 indicators, and obtain the machine learning prediction model.

[0012] S5. Fine-tune the hyperparameters through grid search or random search combined with cross-validation, and select the optimal hyperparameter combination to update the model; use the feature importance evaluation method in random forest to re-evaluate the importance of input features, remove features with small contributions and try new feature combinations; train multiple different neural network models, and integrate the prediction results through weighted averaging to improve accuracy and stability; finally, use the optimized machine learning model to predict the fracture pressure of large-displacement wells in stratified shale.

[0013] Furthermore, the S1 also includes:

[0014] By coring along bedding directions at different angles, vacuuming the cores to remove residual air and maintaining the cores, further experiments on the tensile strength of bedding shale were carried out to obtain the variation characteristics of the tensile strength of bedding shale at different bedding angles, providing data support for the subsequent construction of a bedding model.

[0015] Furthermore, the S2 also includes:

[0016] S21, the rock tensile strength of the layered shale at different bedding angles obtained in S1 is used to optimize the transverse isotropic fracture pressure criterion, where the calculation formula of the single weak plane criterion is as follows:

[0017]

[0018] Where: S t (θ) is the tensile strength under different bedding angles, MPa; θ is the angle between the bedding plane and the horizontal stress, °; T b is the tensile strength of the bedding plane, MPa; T m is the tensile strength of the rock matrix, MPa; θ * is the critical angle, °, given by Sure.

[0019] Based on the above criteria, the calculation of the bursting pressure can be achieved.

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

[0021] Where: P f (θ) is the fracture pressure under different bedding angles, MPa; σ H and σ h are the maximum and minimum horizontal stresses, MPa; α is the Biot coefficient; P p is the formation pore pressure; S t (θ) is the tensile strength at different bedding angles, MPa;

[0022] After the fracture pressure value is calculated, it is usually necessary to calibrate and integrate it with the on-site fracture test data to ensure that the fracture pressure value calculated by the mechanism model is consistent with the actual on-site situation.

[0023] S22. It is necessary to construct a combined spring model of anisotropic geostress field as shown below:

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

[0025]

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

[0027] 2. Generalized Hooke's law expression: Generalized Hooke's law is used to describe the relationship between principal stress and strain in a transversely isotropic elastic body. Its expression is:

[0028]

[0029] In this formula, Among them, σ H , σ h and σ v Respectively represent the maximum horizontal stress, minimum horizontal stress and vertical stress; α is the Biot coefficient; P p is the pore pressure.

[0030] 3. Combined spring model for calculating ground stress: During tectonic movement, the deformation effect of each tectonic movement on the stratum is coordinated and consistent, that is, the deformation of each layer is the same, but due to the difference in stiffness (elastic modulus and Poisson's ratio) of each layer, the tectonic stress generated inside is different. Based on this principle, the σ H and σ h It can be calculated by the following formula:

[0031]

[0032] Furthermore, S3 also includes:

[0033] S31. According to the wellbore structural conditions of the stratified shale extended displacement well, combined with the anisotropic fracture criteria and related physical models constructed above, the fracture pressure calculation of the stratified shale extended displacement well is carried out. The numerical simulation method is used to simulate the stress distribution around the well under different wellbore structural conditions, and then the fracture pressure is determined. During the calculation process, the variation law of the fracture pressure under different wellbore structural parameters (well inclination angle, azimuth) and formation parameters (vertical depth, elastic modulus, Poisson's ratio, ground stress, porosity, tensile strength, pore pressure) is recorded. The obtained fracture pressure value is compared and verified with the actual collected data.

[0034] S32. The accuracy of the calculation model can be evaluated by calculating the average error index (if ME is close to 0, it means that the average deviation of the model prediction value is small and the model accuracy is high. When ME is a positive value, it means that the overall prediction value of the model is too high; when it is a negative value, it means that the overall prediction value of the model is 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, it can be considered that the model has a certain degree of accuracy, but this is not an absolute standard and needs to be analyzed in combination with specific problems). If the error is large, it is necessary to check the calculation model and input parameters, and make necessary corrections and adjustments.

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

[0036]

[0037] Where: n represents the number of samples, that is, the total number of data points involved in calculating the average error; is the model's predicted value for the i-th sample; y i is the true value of the i-th sample.

[0038] S33. Standardize the wellbore structural parameters (well inclination angle, azimuth), formation parameters (vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, pore pressure) and fracture pressure values, and normalize or standardize the input features of different dimensions and numerical ranges. The normalization method can be used to map the data to the 0-1 interval to improve the convergence speed and accuracy of the machine learning algorithm and avoid the dominant influence of certain features on model training due to excessively large or small values.

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

[0040] Furthermore, the S4 also includes:

[0041] S41. Determine the WT-LSTM model structure. First, the number of input layer nodes is determined according to the number of selected input features. Nine features, including well inclination, azimuth, vertical depth, elastic modulus, Poisson's ratio, ground stress, porosity, tensile strength, and pore pressure, are selected, and the number of input layer nodes is 9. In the WT-LSTM model, the input data is first transformed by wavelet transform to decompose it into subsequences of different frequency components. These subsequences are input into the long short-term memory network (LSTM) respectively. The number of output layer nodes is kept at 1, which is used to output the predicted fracture pressure value of the stratified shale.

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

[0043] 1. Determine the number of LSTM hidden layers: Given the characteristics of LSTM in the WT-LSTM model, start the experiment with one LSTM hidden layer. Build the model using the number of nodes preliminarily estimated based on the number of input features and the complexity of the problem. Train the model on the training set and evaluate the performance on the validation set (with root mean square error and accuracy as indicators). If the performance does not meet expectations, gradually increase the number of LSTM hidden layers, repeat the training and evaluation process after each increase, and pay attention to the performance trend of the model on the validation set. Use the L2 regularization method (set the initial regularization coefficient) to control the model complexity when increasing the number of LSTM hidden layers to prevent overfitting. Determine the appropriate range of LSTM hidden layers based on performance and overfitting.

[0044] 2. Determine the number of nodes in the LSTM hidden layer (for the determined number of LSTM hidden layers): For each LSTM hidden layer, determine the initial range of the number of nodes 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 layer of LSTM hidden layer is set to 10-30 (with a step size of 5). If it is determined that there is a second layer of LSTM hidden layer, the initial range of the number of nodes is set to 5-20 (with a step size of 5). Use grid search or random search methods to traverse various combinations within the set range of node numbers, build a model, train it on the training set, and evaluate the performance on the validation set. Observe the changes in model performance under different combinations of node numbers, and select the node number combination with the best performance on the validation set. If the performance difference is small (RMSE difference is considered not obvious within 0.1-0.5 MPa), give priority to combinations with fewer nodes to reduce computational costs and overfitting risks.

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

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

[0047] 2. Use mean square error (MSE) to measure the difference between the model prediction value and the true value, and use the minimization loss function to optimize the model parameters to make the model's prediction results as close to the true value as possible.

[0048] 3. Use the mini-batch stochastic gradient descent (mini-batch SGD) algorithm to update the model parameters of the LSTM part. Divide the data preprocessed by wavelet transform into several small batches (for example, each small batch contains 32 samples), and use a small batch of data each time to calculate the gradient of the LSTM part. Calculate the gradient of the loss function to the LSTM model parameters, and then update the parameters according to the formula.

[0049]

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

[0051] 4. During the training process, use a learning rate decay strategy. For example, use an exponential decay method, set the initial learning rate α0 = 0.01, the decay factor β = 0.95, and at each iteration t, according to the formula α t =α0β t Calculate the current learning rate to gradually reduce the learning rate as training progresses, allowing the model to adjust parameters more finely.

[0052] 5. To prevent overfitting, L2 regularization method is used. Add regularization term to the loss function (where λ is the regularization coefficient and W is the model weight set), the model complexity is limited by penalizing the size of the model weight. At the same time, the Dropout technology can be combined to randomly discard some hidden layer neurons with a certain probability (such as 0.5) during the training process to increase the robustness of the model.

[0053] 6. After completing each epoch (traversing the training set once), calculate the root mean square error (RMSE) and determination coefficient (R 2 ) and other evaluation indicators. Pay close attention to the changes in these indicators. If the RMSE of the validation set does not show a downward or upward trend for several consecutive epochs (5), terminate the training early to prevent overfitting and select the model with the best performance.

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

[0055] Furthermore, the S5 also includes:

[0056] S51. Hyperparameter adjustment: Fine-tune the hyperparameters of the trained model, such as learning rate, regularization coefficient, number of hidden layer nodes, etc. Use grid search or random search combined with cross-validation to search in a wider range of parameter values ​​to find the hyperparameter combination that optimizes model performance. Based on the previously determined initial range of the number of hidden layer nodes, further refine the search step size, search the learning rate in the range of [0.001, 0.1] with a smaller step size (such as 0.001), and search the regularization coefficient in the range of [0.01, 0.1]. Through multiple cross-validations (5-fold cross-validation), evaluate the performance of the model on the validation set under different hyperparameter combinations (such as root mean square error, determination coefficient), and select the hyperparameter combination with the best performance to update the model.

[0057] S52. Feature selection and importance assessment: Re-evaluate the importance of input features and use a feature importance assessment algorithm (feature importance assessment method in random forest). Based on the feature importance score, remove features that contribute less to model prediction and retain key features to reduce model complexity and improve computational efficiency. If it is found that a geological parameter feature has an extremely low importance score for fracture pressure prediction, consider removing it from the input features and then retraining the model to observe performance changes. At the same time, try to combine or transform some features, such as multiplying or adding some related features to obtain new features, and then evaluate the impact of the new feature combination on model performance again, and select a feature combination that can improve model performance.

[0058] S53, Model Integration: Use model integration technology to improve prediction accuracy and stability. Train multiple neural network models with different initial parameters or different structures (such as slightly different numbers of hidden layers), and then integrate the prediction results of these models through weighted averaging. In the weighted averaging method, the weights of each model are determined according to its performance on the validation set, and the model with better performance is given a larger weight. Through model integration, the error of a single model can be effectively reduced and the reliability of the overall prediction can be improved.

[0059] The beneficial effects of the present invention are as follows: by acquiring multi-source data to conduct experiments to construct anisotropic fracture criteria, combining wellbore structural conditions to calculate and verify to form a database, comprehensively considering multiple factors to improve the accuracy of fracture pressure prediction, and providing accurate data support for engineering design; the constructed related physical models make the calculation more scientific and reasonable, and recording and verifying the rules in the calculation helps to ensure accuracy; screening and processing data to generate high-quality data sets, optimizing the model structure and training process, improving the calculation accuracy and efficiency, and enabling the model to better learn features for accurate prediction; accurate prediction helps to reasonably select parameters and control pressure in large-reach well construction, avoid construction problems, and improve adaptability and safety; ultimately providing a basis for engineering decision-making and facilitating resource development. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 Flow chart of a method for predicting fracture pressure of stratified shale extended reach wells based on machine learning model

[0062] Figure 2 The relationship between circumferential stress and wellbore angle under different bedding angles and different tensile strengths

[0063] Figure 3 Schematic diagram of fracture pressure calculation under the conditions of well inclination angle 45° and phase angle 0°

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

[0065] Figure 5 Schematic diagram of fracture pressure calculation under the conditions of well inclination angle 45° and phase angle 90°

[0066] Figure 6 Schematic diagram of the change in formation fracture pressure calculation error under different hidden layer conditions

[0067] Figure 7 This is a comparison table of field ground failure experimental data, mechanism calculation model failure pressure data and machine learning driven calculation model failure pressure under different depth conditions.

Claims

1. A method for predicting fracture pressure of stratified shale extended reach wells based on a machine learning model, characterized in that: include: S1. Obtain the logging and core data of the drilled wells. Take cores along different bedding angles, vacuum the cores to remove residual air and maintain them, then conduct anisotropic tensile strength tests on bedding shale to obtain the variation characteristics of rock tensile strength under different bedding angles, and then construct anisotropic fracture criteria for shale; S2. Based on the rock tensile strength 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, a combined spring model of the anisotropic geostress field of the layered shale is constructed, and the calculated fracture pressure value is corrected and integrated in combination with the field ground failure experimental data to complete the fracture pressure modeling of the layered shale large displacement well; S3. Record the variation patterns of fracture pressure under different wellbore structural parameters (well 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 calculation values ​​with the relevant data of hydraulic fracturing wells in the oil field block, evaluate the accuracy of the calculation model by calculating the average error index, and make corrections and adjustments when necessary based on the evaluation results. Then perform standardization processing, divide the processed data into training sets and test sets according to a certain ratio, and form a fracture pressure database for stratigraphic shale large-displacement wells; S4. Determine the WT-LSTM model structure, set the number of LSTM hidden layers and the number of nodes in each layer, use normal or uniform distribution to initialize the weights and bias parameters of the LSTM part, use mean square error to measure the difference between the predicted value and the true value, use the small 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 indicators, and obtain the machine learning prediction model. S5. Fine-tune the hyperparameters through grid search or random search combined with cross-validation, and select the optimal hyperparameter combination to update the model; use the feature importance evaluation method in random forest to re-evaluate the importance of input features, remove features with small contributions and try new feature combinations; train multiple different neural network models, and integrate the prediction results through weighted averaging to improve accuracy and stability; finally, use the optimized machine learning model to predict the fracture pressure of large-displacement wells in stratified shale.

2. According to the method for predicting fracture pressure of shale extended reach wells based on a machine learning model according to claim 1, said S1 further comprises: By coring along bedding directions at different angles, vacuuming the cores to remove residual air and maintaining the cores, further experiments on the tensile strength of bedding shale were carried out to obtain the variation characteristics of the tensile strength of bedding shale at different bedding angles, providing data support for the subsequent construction of a bedding model.

3. According to the method for predicting fracture pressure of shale extended reach wells based on a machine learning model according to claim 1, said S2 further comprises: S21, the rock tensile strength of the layered shale at different bedding angles obtained in S1 is used to optimize the transverse isotropic fracture pressure criterion, where the calculation formula of the single weak plane criterion is as follows: Where: S t (θ) is the tensile strength under different bedding angles, MPa; θ is the angle between the bedding plane and the horizontal stress, °; T b is the tensile strength of the bedding plane, MPa; T m is the tensile strength of the rock matrix, MPa; θ * is the critical angle, °, given by Sure. Based on the above criteria, the calculation of the bursting pressure can be achieved. P f P(θ) = 3σ h -σ H -αp p +S t P(θ) (Equation 2) Where: P f (θ) is the fracture pressure under different bedding angles, MPa; σ H and σ h are the maximum and minimum horizontal stresses, MPa; α is the Biot coefficient; P p is the formation pore pressure; S t (θ) is the tensile strength at different bedding angles, MPa; After the fracture pressure value is calculated, it is usually necessary to calibrate and integrate it with the on-site fracture test data to ensure that the fracture pressure value calculated by the mechanism model is consistent with the actual on-site situation. S22. It is necessary to construct a combined spring model of the anisotropic geostress field of stratified shale, as shown below: (1) Elastic parameter relationship: For a stratum with a single weak plane (assuming the plane is in the XOY plane), the Poisson's ratio and elastic modulus of anisotropic elastic bodies have specific relationships in different directions, such as: Where: μ H and E H Represent the Poisson's ratio and elastic modulus of the parallel plane, μ ν and E ν are the Poisson's ratio and elastic modulus of the vertical plane, respectively. (2) Generalized Hooke's law expression: Generalized Hooke's law is used to describe the relationship between principal stress and strain in a transversely isotropic elastic body. Its expression is: In this formula, Among them, σ H , σ h and σ v Respectively represent the maximum horizontal stress, minimum horizontal stress and vertical stress; α is the Biot coefficient; P p is the pore pressure. (3) Combined spring model for calculating geostress: During tectonic movement, the deformation effects of various tectonic movements on the strata are coordinated and consistent, that is, the deformation of each layer is the same, but due to the differences in stiffness (elastic modulus and Poisson's ratio) of each layer, the tectonic stress generated inside is different. Based on this principle, the σ H and σ h It can be calculated by the following formula:

4. According to the method for predicting fracture pressure of shale extended reach wells based on a machine learning model according to claim 1, said S3 further comprises: S31. According to the wellbore structural conditions of the stratified shale extended displacement well, combined with the anisotropic fracture criteria and related physical models constructed above, the fracture pressure calculation of the stratified shale extended displacement well is carried out. The numerical simulation method is used to simulate the stress distribution around the well under different wellbore structural conditions, and then the fracture pressure is determined. During the calculation process, the variation law of the fracture pressure under different wellbore structural parameters (well inclination angle, azimuth) and formation parameters (vertical depth, elastic modulus, Poisson's ratio, ground stress, porosity, tensile strength, pore pressure) is recorded. The obtained fracture pressure value is compared and verified with the actual collected data. S32. The accuracy of the calculation model can be evaluated by calculating the average error index (if ME is close to 0, it means that the average deviation of the model prediction value is small and the model accuracy is high. When ME is a positive value, it means that the overall prediction value of the model is too high; when it is a negative value, it means that the overall prediction value of the model is 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, it can be considered that the model has a certain degree of accuracy, but this is not an absolute standard and needs to be analyzed in combination with specific problems). If the error is large, it is necessary to check the calculation model and input parameters, and make necessary corrections and adjustments. The formula for calculating the average error is: Where: n represents the number of samples, that is, the total number of data points involved in calculating the average error; is the model's predicted value for the i-th sample; y i is the true value of the i-th sample. S33. Standardize the wellbore structural parameters (well inclination angle, azimuth), formation parameters (vertical depth, elastic modulus, Poisson's ratio, geostress, porosity, tensile strength, pore pressure) and fracture pressure values, and normalize or standardize the input features of different dimensions and numerical ranges. The normalization method can be used to map the data to the 0-1 interval to improve the convergence speed and accuracy of the machine learning algorithm and avoid the dominant influence of certain features on model training due to excessively large or small values. S34. Divide the processed data into a training set and a test set according to a certain ratio (e.g., 70%-80% for training set and 20%-30% for test set) to generate a data set for subsequent model training and evaluation.

5. According to the method for predicting fracture pressure of shale extended reach wells based on a machine learning model according to claim 1, said S4 further comprises: S41. Determine the structure of the WT-LSTM model: The number of input layer nodes is determined according to the number of selected input features. Nine features are selected, namely, well inclination, azimuth, 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 transformed by wavelet transform to decompose it into subsequences of different frequency components. These subsequences are input into the long short-term memory network (LSTM) respectively. The number of output layer nodes is kept at 1, which is used to output the predicted fracture pressure value of the stratified shale. S42. The hidden layer needs to be set as follows: (1) Determine the number of LSTM hidden layers: Given the characteristics of LSTM in the WT-LSTM model, start the experiment with one LSTM hidden layer. Build the model using the number of nodes preliminarily estimated based on the number of input features and the complexity of the problem. Train the model on the training set and evaluate the performance on the validation set (using root mean square error and accuracy as indicators). If the performance does not meet expectations, gradually increase the number of LSTM hidden layers, repeat the training and evaluation process after each increase, and pay attention to the performance change trend of the model on the validation set. Use the L2 regularization method (set the initial regularization coefficient) to control the model complexity when increasing the number of LSTM hidden layers to prevent overfitting. Determine the appropriate range of LSTM hidden layers based on performance and overfitting conditions. (2) Determination of the number of nodes in the LSTM hidden layer (for the determined number of LSTM hidden layers): For each LSTM hidden layer, determine the initial range of the number of nodes 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 layer of LSTM hidden layer is set to 10-30 (with a step size of 5). If it is determined that there is a second layer of LSTM hidden layer, the initial range of its number of nodes is set to 5-20 (with a step size of 5). Use grid search or random search methods to traverse various combinations within the set range of node numbers, build a model, train it on the training set, and evaluate the performance on the validation set. Observe the changes in model performance under different combinations of node numbers, and select the node number combination with the best performance on the validation set. If the performance difference is small (RMSE difference is considered insignificant within 0.1-0.5 MPa), give priority to combinations with fewer nodes to reduce computational costs and overfitting risks. S43, the specific steps of model training are as follows: (1) The weights and bias parameters of the LSTM part of the WT-LSTM model are initialized using normal distribution or uniform distribution and set to a smaller value (close to 0) to avoid instability caused by excessive gradients in the early stages of training. (2) The mean square error (MSE) is used to measure the difference between the model prediction value and the true value, and the model parameters are optimized by minimizing the loss function to make the model's prediction results as close to the true value as possible. (3) Use the mini-batch stochastic gradient descent (mini-batch SGD) algorithm to update the model parameters of the LSTM part. Divide the data preprocessed by wavelet transform into several small batches (for example, each small batch contains 32 samples), and use a small batch of data each time to calculate the gradient of the LSTM part. Calculate the gradient of the loss function with respect to the LSTM model parameters, and then update the parameters according to the formula. Where: θ is the model parameter, α is the learning rate, m is the number of mini-batch samples, is the gradient of the loss function of the i-th sample in the mini-batch with respect to the model parameters (4) During the training process, use a learning rate decay strategy. For example, use an exponential decay method, set the initial learning rate α0 = 0.01, the decay factor β = 0.95, and at each iteration t, according to the formula α t =α0β t Calculate the current learning rate to gradually reduce the learning rate as training progresses, allowing the model to adjust parameters more finely. (5) To prevent overfitting, the L2 regularization method is used. Add the regularization term to the loss function (where λ is the regularization coefficient and W is the model weight set), the model complexity is limited by penalizing the size of the model weight. At the same time, the Dropout technology can be combined to randomly discard some hidden layer neurons with a certain probability (such as 0.5) during the training process to increase the robustness of the model. (6) After completing each epoch (traversing the training set once), calculate the root mean square error (RMSE) and coefficient of determination (R 2 ) and other evaluation indicators. Pay close attention to the changes in these indicators. If the RMSE of the validation set does not show a downward or upward trend for several consecutive epochs (5), terminate the training early to prevent overfitting and select the model with the best performance. (7) Repeat steps 3-6 until the training stop condition is met (for example, the preset maximum number of training times, such as 200 epochs, or the performance of the validation set no longer improves), and save the best model obtained from the training, including the model structure and parameters, for subsequent prediction applications.

6. According to the method for predicting fracture pressure of shale extended reach wells based on a machine learning model according to claim 1, said S5 further comprises: S51. Hyperparameter adjustment: Fine-tune the hyperparameters of the trained model, such as learning rate, regularization coefficient, number of hidden layer nodes, etc. Use grid search or random search combined with cross-validation to search in a wider range of parameter values ​​to find the hyperparameter combination that optimizes model performance. Based on the previously determined initial range of the number of hidden layer nodes, further refine the search step size, search the learning rate in the range of [0.001, 0.1] with a smaller step size (such as 0.001), and search the regularization coefficient in the range of [0.01, 0.1]. Through multiple cross-validations (5-fold cross-validation), evaluate the performance of the model on the validation set under different hyperparameter combinations (such as root mean square error, determination coefficient), and select the hyperparameter combination with the best performance to update the model. S52. Feature selection and importance assessment: Re-evaluate the importance of input features and use a feature importance assessment algorithm (feature importance assessment method in random forest). Based on the feature importance score, remove features that contribute less to model prediction and retain key features to reduce model complexity and improve computational efficiency. If it is found that a geological parameter feature has an extremely low importance score for fracture pressure prediction, consider removing it from the input features and then retraining the model to observe performance changes. At the same time, try to combine or transform some features, such as multiplying or adding some related features to obtain new features, and then evaluate the impact of the new feature combination on model performance again, and select a feature combination that can improve model performance. S53, Model Integration: Use model integration technology to improve prediction accuracy and stability. Train multiple neural network models with different initial parameters or different structures (such as slightly different numbers of hidden layers), and then integrate the prediction results of these models through weighted averaging. In the weighted averaging method, the weights of each model are determined according to its performance on the validation set, and the model with better performance is given a larger weight. Through model integration, the error of a single model can be effectively reduced and the reliability of the overall prediction can be improved.

Images