Tight sandstone hydraulic fracture crossing prediction method based on machine learning

Abstract

The invention relates to a tight sandstone hydraulic fracture crossing prediction method based on machine learning. Firstly, key parameters of a reservoir and an interlayer are determined through logging data analysis; by collecting actual hydraulic fracturing data, constructing a data set containing stratum parameters and construction parameters, removing abnormal values by using a Z-score method, designing a logging data prediction model based on a recurrent neural network, and generating synthetic data to enhance the integrity of the data set, the problem of logging data information loss is effectively solved, and on the basis, the data prediction method is more accurate and efficient. A decision tree, a support vector machine and logistic regression are adopted as base classifiers, the weight is calculated through an error optimal least square method and a mean square error reciprocal method, and finally a weighted average method is used to obtain an integrated classifier F (x) used for predicting the layer crossing condition of the hydraulic fracture. The method is clear in step, easy to operate and capable of rapidly conducting layer crossing prediction on the hydrofracture horizontal well in a specific block, and a scientific basis is provided for formulating reasonable fracturing parameters.

Description

Technical Field

[0001] The invention relates to a method for predicting hydraulic fracture penetration in dense sandstone based on machine learning, and belongs to the technical field of fracture penetration prediction. Background Art

[0002] In the process of oil and gas extraction, hydraulic fracturing technology is widely used to increase the production of oil and gas wells. The key lies in the accurate prediction and control of hydraulic fractures. However, the penetration behavior of hydraulic fractures is affected by a variety of geological factors. Accurately predicting its penetration is crucial to improving oil well production and optimizing fracturing design. For example, improper fracturing parameter settings may cause the fractures to penetrate the interlayer longitudinally, thereby affecting the fracturing effect. Reasonable control of the longitudinal expansion of the fracture can maximize the extension of the fracture inside the reservoir, increase the reservoir transformation volume, and prevent the penetration of the interlayer to cause the intrusion of water layer fluids, thereby avoiding fracturing failure.

[0003] Current technologies mostly rely on empirical formulas or numerical simulation software, but they usually lack accuracy and versatility. Early studies mainly relied on analysis to predict the expansion of cracks, such as the KGD and PKN models. These models are based on two-dimensional or three-dimensional elastic theory, but usually assume that the material is homogeneous and elastic, ignoring the complexity of actual geological conditions, such as the heterogeneity and anisotropy of rocks.

[0004] In addition, linear elastic fracture mechanics (LEFM) takes into account the stress intensity factor at the crack tip, but due to its assumption based on linear elasticity, it may not accurately describe the expansion of cracks in low permeability coalbed methane reservoirs, which usually exhibit inelastic behaviors such as plasticity or toughness. In order to overcome the limitations of LEFM, scholars have introduced elastic-plastic constitutive models, which can more accurately simulate the nonlinear damage evolution and crack expansion in coal seams. However, these models may require complex mathematical operations and consume more computing resources. In addition, there are a variety of numerical simulation methods, including finite element method (FEM), finite difference method (FDM), boundary element method (BEM), displacement discontinuity method (DDM), discrete element method (DEM) and extended finite element method (XFEM). In addition to the above methods, Kresse proposed an unconventional fracture model (UFM) in 2013 to simulate the high expansion of fractures in natural fracture networks; Yamamoto described a three-dimensional fracture propagation model based on finite element method (FEM) in 2004. Dahi-Taleghani used the extended finite element method (XFM) in 2011 to study the propagation of multiple hydraulic fractures and explored the interaction between induced fractures and natural fractures; Wu proposed a three-dimensional non-planar fracture height extension numerical model based on the displacement discontinuity method (DDM) in 2015; Guo used the cohesive zone model (CZM) in 2017 to study the interaction between hydraulic fractures and interfaces under preset path conditions; Yan used the finite discrete element method (FDEM) in 2017 to construct a three-dimensional fully coupled model considering pore permeability to simulate the penetration behavior of hydraulic fractures. These methods can simulate more complex geological conditions, but their accuracy depends on the selected material parameters and model complexity. For this reason, the present invention is proposed. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method for predicting hydraulic fracture penetration in dense sandstone based on machine learning.

[0006] The technical solution of the present invention is as follows:

[0007] A method for predicting hydraulic fracture penetration in dense sandstone based on machine learning, the steps are as follows:

[0008] S1. Analyze and calculate formation parameters (reservoir and barrier parameters) using logging data, including reservoir thickness, barrier thickness, formation tensile strength and reservoir-barrier stress difference;

[0009] S2. Collect relevant data of horizontal wells that have been hydraulically fractured in the oil field block, including formation parameters that affect hydraulic fracture penetration (reservoir thickness, interlayer thickness, formation tensile strength and reservoir-interlayer stress difference), construction parameters (fracturing fluid displacement, fracturing fluid viscosity) and corresponding penetration results (1 represents penetration, 0 represents no penetration);

[0010] S3, using the Z-score method to remove outliers from the data collected in steps S1 and S2, and due to human factors, instrument failures and other reasons, some well logging data are often distorted or missing in actual applications, resulting in missing formation parameters. Therefore, a comprehensive logging data prediction model based on deep learning is designed to generate synthetic data, increase the integrity of the data set, and generate a data set after standardizing the entire data including the original data and the synthetic data;

[0011] S4, select three base classifiers: decision tree, support vector machine, and logistic regression. For each base classifier, use the Bootstrap sampling method to generate multiple subsets from the training set, and use the genetic algorithm to automatically adjust the parameters of the base classifier;

[0012] S5, averaging the weights calculated by the error optimal least square method and the mean square error inverse method to obtain an average weight;

[0013] S6. The three base classifiers are weighted averaged to obtain the final integrated classifier F(x). The new formation parameters and construction parameters are input into the final integrated classifier F(x) to predict the penetration of hydraulic fractures.

[0014] According to a preferred embodiment of the present invention, in step S1, the formation parameters are calculated as follows:

[0015] S11. By calculating the rock porosity, permeability and formation mud content V sh Determine reservoir thickness and barrier thickness;

[0016] S12, taking the formation tensile strength as the target function and taking the rock density and the P-wave velocity logging data as independent variables to perform function regression to obtain the formation tensile strength;

[0017] S13. Calculate the reservoir-interlayer stress difference using the principal stress value in the reservoir and the principal stress value in the interlayer in the logging data.

[0018] According to the preferred embodiment of the present invention, in step S11, the formation mud content V sh Calculate by formula (2);

[0019]

[0020] Where: I GR is the natural gamma index (dimensionless); GR is the natural gamma logging value (API); GR max is the natural gamma logging value (API) of pure mudstone; GR min is the natural gamma logging value (API) at pure sandstone;

[0021]

[0022] Where: V sh is the formation mud content (dimensionless); I GR is the natural gamma index (dimensionless); GCUR is the Hilch index (dimensionless), which is 3.7 for Tertiary strata and 2.0 for older strata;

[0023] A multivariate regression prediction model was established using the rock porosity measurement value, three rock porosity curves (sonic wave, density, neutron), formation shale content and core porosity. The results showed that the porosity explained by the model was similar to the measured porosity, with a relatively small error.

[0024]

[0025] Where: ρ is the rock density (g / cm 3 ); CNL is the compensation neutron (%); AC is the acoustic time difference (μs / m); is the rock porosity (%); V sh is the formation mud content (dimensionless);

[0026] According to the correlation between rock porosity and permeability, a permeability interpretation model is established. The calculation model is:

[0027] K=0.0036e 0.652φ (4)

[0028] Where: K is permeability (mD); φ is rock porosity (%);

[0029] When V sh If V is less than or equal to 0.2, the rock porosity is greater than or equal to 6% and the permeability is greater than or equal to 0.02 mD, the layer is considered to be a reservoir, and the thickness of the layer is the reservoir thickness; if V sh If the mud content is greater than 0.2, the porosity is less than 6% and the permeability is less than 0.02mD, the layer is considered to be an interlayer, and the thickness of the layer is the thickness of the interlayer.

[0030] Preferably, in step S12, tensile strength is used as the research target, rock density and P-wave velocity logging data are used as independent variables, and a multivariate regression model is established through functional regression analysis to determine the regression equation with the highest correlation:

[0031]

[0032] Where: t is the tensile strength of the formation (MPa); ρ is the density of shale (g / cm 3 );V pis the longitudinal wave velocity (μs / m).

[0033] According to a preferred embodiment of the present invention, in step S13, the reservoir-interlayer stress difference calculation formula is:

[0034] Δσ=σ r -σ i (6)

[0035] Where: Δσ is the reservoir-interlayer stress difference (MPa); σ r is the principal stress value in the reservoir (MPa); σ i is the principal stress value in the interlayer (MPa).

[0036] Preferably, according to the present invention, step S3 is specifically as follows:

[0037] S31. Z-score is a statistical method that measures the degree of dispersion of data points relative to the mean of a data set. It can be used to detect outliers in a data set. The calculation formula of Z-score is as follows:

[0038]

[0039] Where: X is the value of the data point; μ is the mean value of the data set; σ is the standard deviation of the data set;

[0040] Data points with an absolute value of Z-score greater than 3 are outliers;

[0041] S32. Construct a network including a one-dimensional RNN layer, a recursive module and a recursive skip module that introduce Bi-LSTM, an attention mechanism, and an exogenous variable autoregression (ARX) module. First, transform the logging data to be trained into a short-term depth sequence recursively arranged by depth and a long-term depth sequence arranged by recursive skip. The short-term depth sequence and the long-term depth sequence data are input into the RNN layer of the network respectively. After the time step processing, the output sequence is input into the recursive module and the recursive skip module. The ReLU activation function is used in the RNN operation, and the nonlinear activation function used in the recursive module and the recursive skip module is The numbers are Sigmoid and ReLU functions. The RNN layer initially extracts the local dependency and short-term data features of the input logging data. The recursive module and the recursive skip module both add the dropout mechanism. While memorizing the long-term deep sequence data features, it can effectively solve the gradient vanishing phenomenon that is prone to occur during training. This form of skipping neurons that preserves data features can solve the problem of data information forgetting caused by the large depth span of logging data. Then, the network passes several time steps intercepted by ARX and the output of the two recursive modules into the fully connected layer, and estimates the missing logging data in the output in a linear and nonlinear fusion manner.

[0042] S33. Organize the processed data into a data set for subsequent model training:

[0043] The calculation formula for standardization is as follows:

[0044]

[0045] Where: x' is the standardized value; x is the original data; min(x) and max(x) are the minimum and maximum values ​​in the data set, respectively.

[0046] Preferably, according to the present invention, step S4 is specifically as follows:

[0047] S41, data set division: according to the data set processed in step S3, randomly divide it into 70% training set and 30% test set, 70% of the data is used to train the model, and 30% of the data is used to test the prediction ability of the model;

[0048] S42. Define parameters and their optimization range:

[0049] Decision tree: Define the maximum depth of the tree (range: None to 20), the minimum number of split samples (range: 2 to 10), the minimum number of leaf node samples (range: 1 to 5), the splitting criterion (optional: gini or entropy), and the maximum number of features (range: integer, floating point, or None);

[0050] Support vector machine: Define the kernel function type (linear, poly, rbf, sigmoid), regularization parameter (range 0.1 to 100), kernel coefficient (range 0.1 to 10) and degree of polynomial kernel (range 2 to 3);

[0051] Logistic regression: specifies the regularization type (L1, L2, elasticnet, none), regularization parameter (range 0.01 to 100), solver type (newton-cg, lbfgs, liblinear, sag, saga), multi-class strategy (optional ovr or multinomial) and maximum number of iterations (range 100 to 1000);

[0052] S43. Use the Bootstrap sampling method to generate multiple sub-training sets from the training set. For each sub-training set generated by Bootstrap, randomly generate a set of parameter values ​​as the initial population. The chromosome of each individual is encoded as a set of parameter values. Use the base classifier to evaluate the performance of each individual on each sub-training set generated by Bootstrap. The fitness value is based on the accuracy and F1 score of the model on the validation set.

[0053] S44, select the top 20% of the excellent individuals according to the fitness value to enter the next generation, generate new offspring individuals through crossover operation, perform random mutation on the newly generated offspring to increase the diversity of the population, repeat the selection, crossover and mutation steps until the termination condition is met (reaching the maximum number of iterations of 500 times or the improvement of the fitness value is less than 5%), select the individual with the highest fitness from the final population, and its parameter combination is the optimal parameter of the base classifier;

[0054] The calculation formula of accuracy is as follows:

[0055]

[0056] Where: TP is the true positive example, that is, the number of samples correctly predicted by the model as positive; TN is the true negative example, that is, the number of samples correctly predicted by the model as negative; FP is the false positive example, that is, the number of samples incorrectly predicted by the model as positive (first type error); FN is the false negative example,

[0057] That is, the number of samples that the model incorrectly predicts as negative (Type II error);

[0058] The calculation formula of F1 Score is as follows:

[0059]

[0060] Where: Precision refers to the proportion of positive classes predicted by the model to actually be positive classes, and the calculation formula is: Recall refers to the proportion of positive classes that are actually predicted by the model to be positive. The calculation formula is:

[0061]

[0062] Preferably, according to the present invention, step S5 is specifically:

[0063] S51, the error optimal least squares weight optimization calculation formula is as follows:

[0064]

[0065] l 1min +l 2min +l 3min =1(12)

[0066] Where: n is the number of samples in the data set; y i is the actual layer penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1min , l 2min , l 3minare the weights of the three base classifiers under the error optimal least squares method;

[0067] S52, the weight optimization calculation formula of the inverse mean square error method is as follows:

[0068]

[0069] l 1MSE +l 2MSE +l 3MSE =1(15)

[0070] Where n is the number of samples in the data set; y i is the actual layer penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method;

[0071] S53. Calculate the average weight:

[0072]

[0073] In the formula, l 1 , l 2 , l 3 are the average weights of the three base classifiers; l 1min , l 2min , l 3min are the weights of the three base classifiers under the error optimal least squares method; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method.

[0074] Preferably, according to the present invention, in step S6, the weighted average method is:

[0075] S61. For each sample x, calculate the weighted average of the prediction results of all base classifiers. The formula is as follows:

[0076]

[0077] Where: P(x) represents the predicted probability that sample x belongs to hydraulic fracture penetration; α i is the weight of the i-th base classifier; h i (x)

[0078] is the prediction result of the i-th base classifier for sample x;

[0079] S61. According to the weighted average prediction results, the category with the highest prediction probability is selected as the final prediction, that is:

[0080]

[0081] Where: is the final prediction result.

[0082] The beneficial effects of the present invention are:

[0083] The present invention effectively improves the accuracy and reliability of hydraulic fracture penetration prediction by comprehensively utilizing logging data analysis, data preprocessing, base classifier optimization, and integrated learning technology. Specifically, the method first determines the key parameters of the reservoir and interlayer, such as thickness, tensile strength, and stress difference, through logging data analysis; then, by collecting actual hydraulic fracturing data of tight reservoirs, a data set containing formation parameters and construction parameters is constructed, and the Z-score method is used to remove outliers. Aiming at the problem of missing data, a logging data prediction model based on recursive neural network is designed. By combining short-term and long-term depth sequences, synthetic data is generated to enhance the integrity of the data set, which effectively solves the problem of missing logging data information. On this basis, decision tree, support vector machine and logistic regression are used as base classifiers, and the weights are calculated by error optimal least squares method and inverse mean square error method. Finally, the weighted average method + is used to obtain the integrated classifier F(x), which is used to predict the penetration of hydraulic fractures. The implementation steps of the method are clear and easy to operate. It can quickly predict the penetration of hydraulic fractured horizontal wells in specific blocks, providing a scientific basis for formulating reasonable fracturing parameters, which is helpful to increase the production of oil and gas wells, while avoiding fracturing failure caused by fracture penetration, reducing the cost of oilfield testing, and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic diagram of the process flow of the present invention;

[0085] Figure 2 It is a comprehensive logging interpretation diagram for analyzing key layer penetration mechanical parameters of a horizontal well section in an embodiment of the present invention;

[0086] Figure 3 A data diagram showing the relationship between the thickness of the interlayer and the fracture height of a single well in an embodiment of the present invention;

[0087] Figure 4 A data diagram showing the relationship between reservoir thickness and fracture height of a single well in an embodiment of the present invention;

[0088] Figure 5 It is a statistical diagram of the tight sandstone reservoir penetration stress related data in the embodiment of the present invention. DETAILED DESCRIPTION

[0089] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0090] Embodiment 1:

[0091] like Figure 1 As shown, this embodiment provides a method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning, and the steps are as follows:

[0092] S1. Statistical data on formation parameters and construction parameters of 17 wells in the oilfield work area. The logging data is shown in the figure below. Figure 2 As shown, the logging data is used to analyze and calculate the formation parameters (reservoir and barrier parameters), including reservoir thickness, barrier thickness, formation tensile strength and reservoir-barrier stress difference;

[0093] The formation parameters are calculated as follows:

[0094] S11. By calculating the rock porosity, permeability and formation mud content V sh Determine reservoir thickness and barrier thickness;

[0095] Formation mud content V sh Calculate by formula (2);

[0096]

[0097] Where: I GR is the natural gamma index (dimensionless); GR is the natural gamma logging value (API); GR max is the natural gamma logging value (API) of pure mudstone; GR min is the natural gamma logging value (API) at pure sandstone;

[0098]

[0099] Where: V sh is the formation mud content (dimensionless); I GR is the natural gamma index (dimensionless); GCUR is the Hilch index (dimensionless), which is 3.7 for Tertiary strata and 2.0 for older strata;

[0100] The rock porosity is obtained using logging interpretation data, and the calculation formula is as follows:

[0101] TOC=40.6-15.6*DEN+0.013*GR+0.009*CNL (19)

[0102] φ=0.32*AC-0.55*DEN+0.84*TOC-0.48*I GR +8.67 (20)

[0103] Where: I GRis the natural gamma index (dimensionless); TOC is the organic carbon content (dimensionless); DEN is the rock density (g / cm 3 ); GR is the estimated well section natural gamma logging value (API); CNL is the compensated neutron (%); AC is the acoustic wave delay (μs / m); φ is the rock porosity (%);

[0104] According to the correlation between rock porosity and permeability, a permeability interpretation model is established. The calculation model is:

[0105] K=0.0036e 0.652φ (4)

[0106] Where: K is permeability (mD); φ is rock porosity (%);

[0107] When V sh If V is less than or equal to 0.2, the rock porosity is greater than or equal to 6% and the permeability is greater than or equal to 0.02 mD, the layer is considered to be a reservoir, and the thickness of the layer is the reservoir thickness; if V sh If the shale content is greater than 0.2, the porosity is less than 6% and the permeability is less than 0.02mD, the layer is considered to be an interlayer, and the thickness of the layer is the thickness of the interlayer;

[0108] S12. The calculation formula of formation tensile strength is:

[0109] σ t =14.91ln(V p )+22.41ln(ρ)-140.2 (5)

[0110] Where: t is the tensile strength of the formation (MPa); ρ is the density of shale (g / cm 3 );V p is the longitudinal wave velocity (μs / m).

[0111] S13, calculating the reservoir-interlayer stress difference using the principal stress value in the reservoir and the principal stress value in the interlayer in the logging data;

[0112] The calculation formula of reservoir-interlayer stress difference is:

[0113] Δσ=σ r -σ i (6)

[0114] Where: Δσ is the reservoir-interlayer stress difference (MPa); σ r is the principal stress value in the reservoir (MPa); σ i is the principal stress value in the interlayer (MPa).

[0115] S2. Collect relevant data of horizontal wells that have been hydraulically fractured in the oil field block, including formation parameters that affect hydraulic fracture penetration (reservoir thickness, interlayer thickness, formation tensile strength and reservoir-interlayer stress difference), construction parameters (fracturing fluid displacement, fracturing fluid viscosity) and corresponding penetration results (1 represents penetration, 0 represents no penetration);

[0116] S3, the data collected in steps S1 and S2 are removed from outliers using the Z-score method, and due to human factors, instrument failures and other reasons, some well logging data are often distorted or missing in actual applications, resulting in missing formation parameters. Therefore, a comprehensive logging data prediction model based on deep learning is designed to generate synthetic data, increase the integrity of the data set, and generate a data set after standardizing the entire data including the original data and the synthetic data. Specifically;

[0117] S31. The calculation formula of Z-score is as follows:

[0118]

[0119] Where: X is the value of the data point; μ is the mean value of the data set; σ is the standard deviation of the data set;

[0120] Data points with an absolute value of Z-score greater than 3 are outliers;

[0121] S32. Construct a network including a one-dimensional RNN layer, a recursive module and a recursive skip module that introduce Bi-LSTM, an attention mechanism, and an exogenous variable autoregression (ARX) module. First, transform the logging data to be trained into a short-term depth sequence recursively arranged by depth and a long-term depth sequence arranged by recursive skip. The short-term depth sequence and the long-term depth sequence data are input into the RNN layer of the network respectively. After the time step processing, the output sequence is input into the recursive module and the recursive skip module. The ReLU activation function is used in the RNN operation, and the nonlinear activation function used in the recursive module and the recursive skip module is The numbers are Sigmoid and ReLU functions. The RNN layer initially extracts the local dependency and short-term data features of the input logging data. The recursive module and the recursive skip module both add the dropout mechanism. While memorizing the long-term deep sequence data features, it can effectively solve the gradient vanishing phenomenon that is prone to occur during training. This form of skipping neurons that preserves data features can solve the problem of data information forgetting caused by the large depth span of logging data. Then, the network passes several time steps intercepted by ARX and the output of the two recursive modules into the fully connected layer, and estimates the missing logging data in the output in a linear and nonlinear fusion manner.

[0122] S33. Organize the processed data into a data set for subsequent model training:

[0123] The calculation formula for standardization is as follows:

[0124]

[0125] Where: x' is the standardized value; x is the original data; min(x) and max(x) are the minimum and maximum values ​​in the data set respectively;

[0126] S4. Select three base classifiers: decision tree, support vector machine, and logistic regression. For each base classifier, use the Bootstrap sampling method to generate multiple subsets from the training set, and use the genetic algorithm to automatically adjust the parameters of the base classifier. Specifically;

[0127] S41, data set division: according to the data set processed in step S3, randomly divide it into 70% training set and 30% test set, 70% of the data is used to train the model, and 30% of the data is used to test the prediction ability of the model;

[0128] S42. Define parameters and their optimization range:

[0129] Decision tree: Define the maximum depth of the tree (range: None to 20), the minimum number of split samples (range: 2 to 10), the minimum number of leaf node samples (range: 1 to 5), the splitting criterion (optional: gini or entropy), and the maximum number of features (range: integer, floating point, or None);

[0130] Support vector machine: Define the kernel function type (linear, poly, rbf, sigmoid), regularization parameter (range 0.1 to 100), kernel coefficient (range 0.1 to 10) and degree of polynomial kernel (range 2 to 3);

[0131] Logistic regression: specifies the regularization type (L1, L2, elasticnet, none), regularization parameter (range 0.01 to 100), solver type (newton-cg, lbfgs, liblinear, sag, saga), multi-class strategy (optional ovr or multinomial) and maximum number of iterations (range 100 to 1000);

[0132] S43. Use the Bootstrap sampling method to generate multiple sub-training sets from the training set. For each sub-training set generated by Bootstrap, randomly generate a set of parameter values ​​as the initial population. The chromosome of each individual is encoded as a set of parameter values. Use the base classifier to evaluate the performance of each individual on each sub-training set generated by Bootstrap. The fitness value is based on the accuracy and F1 score of the model on the validation set.

[0133] S44, select the top 20% of the excellent individuals according to the fitness value to enter the next generation, generate new offspring individuals through crossover operation, perform random mutation on the newly generated offspring to increase the diversity of the population, repeat the selection, crossover and mutation steps until the termination condition is met (reaching the maximum number of iterations of 500 times or the improvement of the fitness value is less than 5%), select the individual with the highest fitness from the final population, and its parameter combination is the optimal parameter of the base classifier;

[0134] The calculation formula of accuracy is as follows:

[0135]

[0136] Where: TP is the true positive example, that is, the number of samples correctly predicted by the model as positive; TN is the true negative example, that is, the number of samples correctly predicted by the model as negative; FP is the false positive example, that is, the number of samples incorrectly predicted by the model as positive (first type error); FN is the false negative example, that is, the number of samples incorrectly predicted by the model as negative (second type error);

[0137] The calculation formula of F1 Score is as follows:

[0138]

[0139] Where: Precision refers to the proportion of positive classes predicted by the model to actually be positive classes, and the calculation formula is: Recall refers to the proportion of positive classes that are actually predicted by the model to be positive. The calculation formula is:

[0140] S5. averaging the weights calculated by the error optimal least squares method and the mean square error inverse method to obtain an average weight, specifically;

[0141] S51, the error optimal least squares weight optimization calculation formula is as follows:

[0142]

[0143] l 1min +l 2min +l 3min =1 (12)

[0144] Where: n is the number of samples in the data set; y i is the actual penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1min , l 2min , l 3minare the weights of the three base classifiers under the error optimal least squares method;

[0145] S52, the weight optimization calculation formula of the inverse mean square error method is as follows:

[0146]

[0147] l 1MSE +l 2MSE +l 3MSE =1 (15)

[0148] Where n is the number of samples in the data set; y i is the actual penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method;

[0149] S53. Calculate the average weight:

[0150]

[0151] In the formula, l 1 , l 2 , l 3 are the average weights of the three base classifiers; l 1min , l 2min , l 3min are the weights of the three base classifiers under the error optimal least squares method; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method.

[0152] S6. Take the weighted average of the three base classifiers to obtain the final integrated classifier F(x);

[0153] The weighted average method is:

[0154] S61. For each sample x, calculate the weighted average of the prediction results of all base classifiers. The formula is as follows:

[0155]

[0156] Where: P(x) represents the predicted probability that sample x belongs to hydraulic fracture penetration; α i is the weight of the i-th base classifier; h i (x) is the prediction result of the i-th base classifier for sample x;

[0157] S61. According to the weighted average prediction results, the category with the highest prediction probability is selected as the final prediction, that is:

[0158]

[0159] Where: is the final prediction result.

[0160] Through the reservoir thickness, interlayer thickness, formation tensile strength, reservoir-interlayer stress difference, fracturing fluid displacement, and fracturing fluid viscosity data of the five wells X1, X2, X3, X4, and X5, the final integrated classifier F(x) is used to predict whether the hydraulic fractures of these five wells penetrate the layers. The above is only a partial preferred embodiment of the present invention. Any technician familiar with the field may modify the technical solution described above or modify it into an equivalent technical solution. Therefore, the corresponding simple modifications or equivalent transformations made according to the technical solution of the present invention are within the scope of protection required by the present invention.

Claims

1. A method for predicting hydraulic fracture penetration in dense sandstone based on machine learning, characterized in that: Here are the steps: S1. Analyze and calculate formation parameters using logging data, including reservoir thickness, interlayer thickness, formation tensile strength, and reservoir-interlayer stress difference; S2. Collect relevant data of horizontal wells that have been hydraulically fractured, including formation parameters that affect hydraulic fracture penetration, construction parameters and corresponding penetration results; S3, using the Z-score method to remove outliers from the data collected in steps S1 and S2, designing a comprehensive logging data prediction model based on deep learning to generate synthetic data, increasing the integrity of the data set, and generating a data set after standardizing the entire data including the original data and the synthetic data; S4, select three base classifiers: decision tree, support vector machine, and logistic regression. For each base classifier, use the Bootstrap sampling method to generate multiple subsets from the training set, and use the genetic algorithm to automatically adjust the parameters of the base classifier; S5, averaging the weights calculated by the error optimal least square method and the mean square error inverse method to obtain an average weight; S6. The three base classifiers are weighted averaged to obtain the final integrated classifier F(x). The new formation parameters and construction parameters are input into the final integrated classifier F(x) to predict the penetration of hydraulic fractures.

2. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 1, characterized in that: In step S1, the formation parameters are calculated as follows: S11. By calculating the rock porosity, permeability and formation mud content V sh Determine reservoir thickness and barrier thickness; S12, taking the formation tensile strength as the target function and taking the rock density and the P-wave velocity logging data as independent variables to perform function regression to obtain the formation tensile strength; S13. Calculate the reservoir-interlayer stress difference using the principal stress value in the reservoir and the principal stress value in the interlayer in the logging data.

3. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 2, characterized in that: In step S11, the formation mud content V sh Calculate by formula (2); Where: I GR is the natural gamma index; GR is the natural gamma logging value; GR max is the natural gamma logging value at pure mudstone; GR min is the natural gamma logging value at pure sandstone; Where: V sh I is the formation mud content; GR is the natural gamma index; GCUR is the Hilch index; Where: ρ is rock density; CNL is compensation neutron; AC is acoustic time difference; is the rock porosity; V sh is the formation mud content; According to the correlation between rock porosity and permeability, a permeability interpretation model is established. The calculation model is: Where: K is permeability (mD); φ is rock porosity (%); When V sh If V is less than or equal to 0.2, the rock porosity is greater than or equal to 6% and the permeability is greater than or equal to 0.02 mD, the layer is considered to be a reservoir, and the thickness of the layer is the reservoir thickness; if V sh If the mud content is greater than 0.2, the porosity is less than 6% and the permeability is less than 0.02mD, the layer is considered to be an interlayer, and the thickness of the layer is the thickness of the interlayer.

4. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 3, characterized in that: In step S12, tensile strength is used as the research target, rock density and P-wave velocity logging data are used as independent variables, and a multivariate regression model is established through functional regression analysis: Where: t is the tensile strength of the formation; ρ is the density of shale; V p is the longitudinal wave velocity.

5. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 4, characterized in that: In step S13, the reservoir-interlayer stress difference calculation formula is: Ds = s r -s i (6) Where: Δσ is the stress difference between the reservoir and the barrier layer; σ r is the principal stress value in the reservoir; σ i is the principal stress value in the interlayer.

6. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 5, characterized in that: Step S3 is specifically as follows: S31. The calculation formula of Z-score is as follows: Where: X is the value of the data point; μ is the mean value of the data set; σ is the standard deviation of the data set; Data points with an absolute value of Z-score greater than 3 are outliers; S32. Construct a network including a one-dimensional RNN layer, a recursive module and a recursive skip module that introduce Bi-LSTM, an attention mechanism, and an exogenous variable autoregression (ARX) module. First, transform the logging data to be trained into a short-term depth sequence recursively arranged by depth and a long-term depth sequence arranged by recursive skip. Input the short-term depth sequence and the long-term depth sequence data into the RNN layer of the network respectively. After the time step processing, input the output sequence into the recursive module and the recursive skip module. The ReLU activation function is used in the RNN operation, and the nonlinear activation used in the recursive module and the recursive skip module is used. The activation functions are Sigmoid and ReLU functions. The RNN layer initially extracts the local dependency and short-term data features of the input logging data. The recursive module and the recursive skip module both add the dropout mechanism. While memorizing the features of long-term deep sequence data, they can solve the gradient vanishing phenomenon that is prone to occur during training. This form of skipping neurons that preserves data features solves the problem of data information forgetting caused by the large depth span of logging data. Then, the network passes several time steps intercepted by ARX and the outputs of the two recursive modules into the fully connected layer, and estimates the missing logging data in the output in a linear and nonlinear fusion manner. S33. Organize the processed data into a data set for subsequent model training: The calculation formula for standardization is as follows: Where: x' is the standardized value; x is the original data; min(x) and max(x) are the minimum and maximum values ​​in the data set, respectively.

7. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 6, characterized in that: Step S4 is specifically as follows: S41, data set division: according to the data set processed in step S3, randomly divide it into 70% training set and 30% test set, 70% of the data is used to train the model, and 30% of the data is used to test the prediction ability of the model; S42. Define parameters and their optimization range: Decision tree: Define the maximum depth of the tree, the minimum number of split samples, the minimum number of leaf node samples, the splitting criterion and the maximum number of features; Support vector machine: Define the kernel function type, regularization parameter, kernel coefficient and degree of polynomial kernel; Logistic regression: specify the regularization type, regularization parameter, solver type, multi-class strategy, and maximum number of iterations; S43. Use the Bootstrap sampling method to generate multiple sub-training sets from the training set. For each sub-training set generated by Bootstrap, randomly generate a set of parameter values ​​as the initial population. The chromosome of each individual is encoded as a set of parameter values. Use the base classifier to evaluate the performance of each individual on each sub-training set generated by Bootstrap. The fitness value is based on the accuracy and F1 score of the model on the validation set. S44, select the top 20% of the best individuals according to the fitness value to enter the next generation, generate new offspring individuals through crossover operation, perform random mutation on the newly generated offspring to increase the diversity of the population, repeat the selection, crossover and mutation steps until the termination condition is met, and select the individual with the highest fitness from the final population, and its parameter combination is the optimal parameter of the base classifier; The calculation formula of accuracy is as follows: Where: TP is the true positive example, that is, the number of samples correctly predicted by the model as positive; TN is the true negative example, that is, the number of samples correctly predicted by the model as negative; FP is the false positive example, that is, the number of samples incorrectly predicted by the model as positive; FN is the false negative example, that is, the number of samples incorrectly predicted by the model as negative. The calculation formula of F1 score is as follows: Where: Precision refers to the proportion of positive classes predicted by the model to actually be positive classes, and the calculation formula is: Recall refers to the proportion of positive classes that are actually predicted by the model to be positive. The calculation formula is:

8. The method for predicting the penetration of hydraulic fractures in dense sandstone based on machine learning according to claim 7, characterized in that: Step S5 is as follows: S51. The error optimal least squares weight optimization calculation formula is as follows: l 1min +l 2min +l 3min ＝1(12) Where: n is the number of samples in the data set; y i is the actual penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1min , l 2min , l 3min are the weights of the three base classifiers under the error optimal least squares method; S52, the weight optimization calculation formula of the inverse mean square error method is as follows: l 1MSE +l 2MSE +l 3MSE ＝1(15) Where n is the number of samples in the data set; y i is the actual penetration result; are the predicted values ​​of the three base classifiers: decision tree, support vector machine, and logistic regression; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method; S53. Calculate the average weight: Where l1, l2, and l3 are the average weights of the three base classifiers respectively; l 1min , l 2min , l 3min are the weights of the three base classifiers under the error optimal least squares method; l 1MSE , l 2MSE , l 3MSE are the weights of the three base classifiers under the inverse mean square error method.

9. The method for predicting penetration of tight sandstone hydraulic fractures based on machine learning according to claim 8, characterized in that: In step S6, the weighted average method is: S61. For each sample x, calculate the weighted average of the prediction results of all base classifiers. The formula is as follows: Where: P(x) represents the predicted probability that sample x belongs to hydraulic fracture penetration; α i is the weight of the i-th base classifier; h i (x) is the prediction result of the i-th base classifier for sample x; S61. According to the weighted average prediction results, the category with the highest prediction probability is selected as the final prediction, that is: Where: is the final prediction result.

Images