Shale reservoir brittleness index earthquake intelligent prediction method based on improved BP neural network

Abstract

The invention discloses a shale reservoir brittleness index earthquake intelligent prediction method based on an improved BP neural network, and relates to the technical field of shale reservoir brittleness index prediction methods. According to the method, an improved BP neural network method is adopted to establish a nonlinear relation model between the brittleness index and the elastic parameter, a shale reservoir elastic parameter data body is obtained based on a pre-stack inversion method, and the inversion body is input into the model to obtain a prediction result of the brittleness index of the shale reservoir. According to the invention, the accuracy and stability of model prediction can be improved, and the brittleness index prediction precision is improved.

Description

Technical Field

[0001] The invention relates to the technical field of shale reservoir brittleness index prediction methods, in particular to a shale reservoir brittleness index seismic intelligent prediction method based on an improved BP neural network. Background Art

[0002] Prestack simultaneous inversion is the most widely used inversion method, and its theoretical basis is the Zoeppritz equation and its simplified expression. This inversion technology uses prestack gather data, logging curves and seismic layer data for inversion to obtain P-wave, S-wave and density data. BP neural network was proposed by Rumelhart et al. in 1986 and is one of the most widely used neural network models. The BP neural network model structure includes an input layer, an output layer and any number of hidden layers, each of which is composed of several neurons. During the training process, the model follows the principle of reducing the error between the target output and the actual output, and continuously updates the weights w and thresholds b of the neurons in each hidden layer and the input layer by back propagating the output layer error, so it is called the "error back propagation algorithm", that is, the BP algorithm (Back Propagation).

[0003] The commonly used seismic prediction method for shale brittleness index currently uses multiple pre-stack seismic data volumes for joint inversion to obtain elastic parameter data volumes such as Young's modulus and Poisson's ratio. The elastic parameter data volume is then calculated through linear mathematical relationships or converted using a nonlinear multivariate regression model to directly or indirectly obtain the brittleness index data volume, thereby obtaining the three-dimensional lateral and vertical brittleness index distribution of the shale gas reservoir.

[0004] However, during the research and practice of the above method, the inventors of the present invention found that the mathematical relationship conversion formula between elastic parameters and brittleness index is based on a simple linear mathematical relationship conversion of only a few of the elastic parameters, and the multivariate regression model is easily affected by factors such as model parameter selection and error checking. The accuracy of the brittleness index prediction results needs to be further improved.

[0005] Therefore, a new solution to the above problems needs to be proposed. Summary of the invention

[0006] The purpose of the present invention is to provide a shale reservoir brittleness index seismic intelligent prediction method based on an improved BP neural network, which can improve the seismic prediction accuracy of the brittleness index, thereby more accurately identifying the distribution of favorable areas for shale reservoir fracturing, so as to solve the technical problems raised in the background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a seismic prediction method for shale reservoir brittleness index based on an improved BP neural network, comprising the following steps:

[0008] S1: Obtain logging data and experimental data of several known wells, wherein the logging data are logging parameters at n different depths in the well, including longitudinal wave velocity Vp, shear wave velocity Vs and density ρ, and the experimental data are various rock parameters, including Young's modulus, Poisson's ratio, shear modulus, Lame coefficient and bulk modulus. The rock dynamic parameters calculated by fitting and correcting the logging data using the experimental data are static rock parameters, providing sufficient data points for neural network model training, and randomly selecting one well as a verification well, and the rest as training wells;

[0009] S2: Correlation analysis and comparison are performed on each rock parameter in the training well and the mineral brittleness index. The three rock parameters with the top three absolute values ​​of the correlation coefficient are selected as independent variables, marked as independent variable A, independent variable B, and independent variable C respectively, and the mineral brittleness index is used as the dependent variable;

[0010] S3: Construct the model data set, and use the three independent variables at the same depth of all wells to form sample X i ={X Ai , X Bi , X Ci}, i represents the i-th depth, i = 1, 2, 3, ..., n, X Ai , X Bi , X Ci They represent the values ​​of independent variable A, independent variable B, and independent variable C at the i-th depth respectively. The brittleness index corresponding to this depth is Y i , Y i As the label of the sample, the sample data and the sample label constitute the model data set;

[0011] S4: Establish a training matrix, and construct all sample data into a matrix X = {X A , X B , X C}, where X A , X B , X C For all X Ai , X Bi , X Ci The column vector consists of the brittleness index Y corresponding to the depth i Construct matrix Y; preset an iteration matrix X k ={X Ak , X Bk , X Ck}, the number of iterations N, k = 1 ~ N; and when k = 1, X Ak =X A , X Bk =X B , X Ck =X C ;

[0012] S5: Use trial and error method to calculate the error of different hidden layer node models, determine the number of hidden layer neurons of BP neural network model, and determine the structure of BP neural network model;

[0013] S6: Improve the learning rate in the gradient descent method and use the improved BP neural network model to obtain the optimal parameters for model training;

[0014] S7: Select a verification well, obtain its corresponding data in the model data set, input it into the improved BP neural network brittleness index prediction model, output its predicted value and compare it with the true value, and use the determination coefficient evaluation index to verify the model stability;

[0015] S8: Using the prestack inversion method, the three rock parameter data volumes in S2 are calculated;

[0016] S9: Based on the improved BP neural network model, the rock parameter data body calculated in S8 is used as the neural network input data to predict the shale reservoir brittleness index data body, and then determine the shale reservoir brittleness index seismic prediction results.

[0017] Further, in said S1, the rock parameters Young's modulus (E), Poisson's ratio (μ), bulk modulus (K), shear modulus (G) and Lame coefficient (λ) are calculated using the following formulas;

[0018]

[0019] V p is the longitudinal wave velocity, V s is the shear wave velocity.

[0020] Further, the S6 at least includes the following steps:

[0021] S61: normalize the elements in the matrix X respectively to obtain a normalized matrix X';

[0022] S62: For matrix X k The elements in are normalized respectively to obtain the normalized matrix X k ';

[0023] S63: normalizing the elements in the matrix Y respectively to obtain a normalized matrix Y';

[0024] S64: Calculate the matrix X k 'Predicted value;

[0025] Further, the S64 at least includes the following steps:

[0026] S641: Use the Sigmoid activation function to transform the matrix Xk 'Input into the hidden layer of the neural network to obtain the hidden layer output matrix X KH ';

[0027] S642: Output the hidden layer matrix X KH 'Use the Purelin activation function to input to the output layer and get the output result Y k ;

[0028] S643: Calculate Y k The error between and Y' is denoted as the error function J(θ);

[0029] S644: Determine the gradient ▽J(θ k );

[0030] S645: Using learning rate α k Multiply by the gradient of the error function to get the distance α of the gradient descent when the current iteration number is k k *▽J(θ k );

[0031] S646: When the number of iterations is k, the model parameter θ k For the updated model training parameters, determine that the value of the error function J(θ) is less than the iteration termination condition ε, then the model training ends, otherwise proceed to step S647;

[0032] S647: Update learning rate α k After the update is completed, proceed to step S644.

[0033] Furthermore, the error function J(θ) and the model parameter θ in S64 are k and learning rate α k The specific calculation is based on the following formula:

[0034] J(θ)=||Y k (θ)-Y'|| 2 2

[0035]

[0036] θ is the weight of the input data, Y k (θ) is the predicted output result, Y' is the actual output of the input data, α k is the learning rate when the number of iterations is k, is the gradient of the error function when the number of iterations is k, α is the learning rate, k is the number of iterations, α 0 is the initial learning rate, a constant, taken as 0.5.

[0037] Furthermore, in S641-S643, the normalization process is performed using the following formula:

[0038]

[0039] Z i is the standardized value of each elastic parameter, dimensionless; X i is the original value of each elastic parameter; max(X i )、min(X i ) are the maximum and minimum values ​​of each elastic parameter in the target layer.

[0040] Furthermore, the S7 determination coefficient is calculated using the following formula:

[0041]

[0042] n is the number of samples, is the predicted value of the i-th sample, y i is the true value of the i-th sample, is the average value among the sample prediction values, is the mean of the true values ​​of the samples.

[0043] Further, the S8 at least includes the following steps:

[0044] S81: Super gather processing is performed on the seismic data, and further converted into pre-stack angle gathers;

[0045] S82: Extract seismic wavelets and complete the calibration of strong seismic reflection interface and logging curve;

[0046] S83: Based on the well location information, layer information and seismic wavelet, establish the initial model of P-wave velocity, S-wave velocity and density required for inversion;

[0047] S84: Use the Zoeppritz equation to perform prestack inversion and obtain the inverted P-wave velocity, S-wave velocity and density data volume;

[0048] S85: Calculate the rock parameter data using rock mechanics formulas.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention adopts an improved BP neural network method to establish a nonlinear relationship model between the brittleness index and the elastic parameters, and obtains the shale reservoir elastic parameter data body based on the prestack inversion method, and obtains the prediction result of the shale reservoir brittleness index by inputting the inversion body into the model;

[0051] The improved BP neural network model construction process includes:

[0052] Preprocess the data set; use the trial and error method to determine the model structure; improve the learning rate in the gradient descent method to an adaptive learning rate that gradually decreases with the number of training iterations; use the improved BP neural network model to obtain the optimal parameters for model training; use the posterior data to evaluate the model and verify the model stability;

[0053] The prestack inversion process includes:

[0054] Prestack seismic data optimization; Well seismic fine calibration; Elastic parameter volume calculation;

[0055] The present invention can improve the accuracy and stability of model prediction and enhance the prediction precision of brittleness index. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0057] Figure 1 It is a flow chart of a method for seismic intelligent prediction of shale reservoir brittleness index based on an improved BP neural network of the present invention;

[0058] Figure 2 It is a schematic diagram of the BP neural network structure of the present invention;

[0059] Figure 3 It is the dynamic and static Young's modulus fitting diagram of the present invention;

[0060] Figure 4 is a correlation analysis diagram of elastic parameters of the present invention;

[0061] Figure 4 (a) Cross plot of brittleness index and Young's modulus; Figure 4 (b) Cross plot of brittleness index and Poisson’s ratio; Figure 4 (c) Cross-plot of brittleness index and shear modulus; Figure 4 (d) Cross plot of brittleness index and bulk modulus; Figure 4 (e) Cross plot of brittleness index and Lame coefficient;

[0062] Figure 5 It is the error analysis diagram of different hidden layer node numbers of the present invention;

[0063] Figure 6 It is a comparison diagram of the error convergence curve of the BP neural network of the present invention and the error convergence curve of the improved BP neural network;

[0064] Figure 7This is a correlation analysis diagram of the model training, test data and real data of the present invention;

[0065] Figure 7 (a) Correlation analysis diagram between training data prediction results and actual data; Figure 7 (b) Correlation analysis diagram between test data prediction results and actual data;

[0066] Figure 8 This is an analysis diagram of the effect verification of the improved BP neural network model of the present invention;

[0067] Figure 8 (a) Correlation analysis diagram between the predicted value and the true value of the single well brittleness index; Figure 8 (b) Comparison of predicted and actual values ​​of the brittleness index of a single well;

[0068] Fig. 9 It is the pre-stack seismic data optimization and well-seismic fine calibration analysis diagram of the present invention;

[0069] Fig. 9 (a) CRP gathers; Fig. 9 (b) Super CDP gathers; Fig. 9 (c) AVA gathers; Fig. 9 (d) Synthetic seismic record of well Z1;

[0070] Fig.10 It is a plane diagram of the pre-stack simultaneous inversion elastic parameter results of the present invention;

[0071] Fig.11 It is a plane distribution diagram of the brittleness index prediction results of the improved BP neural network method of the present invention;

[0072] Fig.11 (a) Plane diagram of Young’s modulus data volume; Fig.11 (b) Plane diagram of Poisson’s ratio data volume; Fig.11 (c) Plane diagram of shear modulus data volume;

[0073] Fig.12 It is a plane comparison diagram of the brittleness index prediction results of the present invention;

[0074] Fig.13 It is a comparison diagram of the predicted value and measured value of shale brittle earthquake based on the improved BP neural network and the Rickman method of the present invention;

[0075] Fig.14 This is a data comparison and analysis chart of the brittleness index prediction results;

[0076] Fig.14 (a) Scatter distribution diagram of predicted and true values ​​of brittleness index; Fig.14(b) Scatter distribution diagram of relative error of brittleness index based on improved BP neural network and Rickman method. DETAILED DESCRIPTION

[0077] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0078] The embodiment of the present invention provides a seismic intelligent prediction method for shale reservoir brittleness index based on an improved BP neural network, which can predict the brittleness index of shale reservoirs, thereby guiding the selection of reservoir volume fracturing transformation areas. Detailed descriptions are given below.

[0079] The technical solution of the present invention includes three main steps: construction of an improved BP neural network model, inversion of prestack seismic elastic parameters, and seismic prediction and comparative analysis of shale reservoir brittleness index based on the improved BP neural network. Figure 1 The present invention is a flow chart of a method for seismic intelligent prediction of shale reservoir brittleness index based on an improved BP neural network. Figure 2 It is a schematic diagram of the BP neural network structure of the present invention.

[0080] The construction process of the improved BP neural network model includes: data set preprocessing; using the trial and error method to determine the model structure; improving the learning rate in the gradient descent method to an adaptive learning rate that gradually decreases with the number of training iterations; using the improved BP neural network model to obtain the optimal parameters for model training; and using the posterior data to evaluate the model to verify the model stability.

[0081] The present invention proposes to use a neural network to predict the brittleness index, which is to seek a mapping relationship between the brittleness index and multiple elastic parameters.

[0082] Therefore, based on the logging data and rock mechanics experimental data of a block in western Sichuan Basin, the present invention constructs the original data set of the model through dynamic and static elastic parameter conversion and optimization of static elastic parameters. Static parameters are rock parameters such as Young's modulus and Poisson's ratio obtained through indoor rock fracture experiments, i.e. triaxial rock mechanics experiments; dynamic parameters are rock elastic parameters calculated using logging data. Rock static parameters have the advantages of high accuracy and reliable data, but the experimental data points are limited and cannot meet the requirements of a large amount of training data for neural networks. Therefore, a large number of static elastic parameter data points can be obtained by correcting continuous dynamic elastic parameters.

[0083] The static elastic parameter data of shale were obtained by triaxial rock mechanics experiments in a well in a block in western Sichuan Basin. The dynamic elastic parameter data of rock calculated by cross-dipole acoustic logging data were fitted to obtain the conversion formula of dynamic and static parameters of rock, thereby realizing the static correction of dynamic parameters and providing sufficient data points for neural network model training.

[0084] The dynamic and static Young's modulus of rock have a linear relationship, but the relationship between dynamic and static Poisson's ratio is unclear. Using depth as a connection, the static Young's modulus obtained from the triaxial experiment in the study area and the dynamic Young's parameters of rock calculated from the cross-dipole acoustic logging data are fitted. The required static elastic parameter data points for model training are further calculated through the fitting formula. Figure 3 It is the dynamic and static Young's modulus fitting diagram of the present invention.

[0085] Exploring the relationship between the model characteristic variables, namely the static elastic parameters, and the target variable, the mineral brittleness index, can further construct the original data set of the model. Perform a brittleness index correlation analysis on the elastic parameters, construct a cross plot of the brittleness index and the elastic parameters, and select the sensitive elastic parameters of the brittleness index based on the positive and negative values ​​of the slope of the fitting equation in the cross plot. A positive slope indicates a positive correlation; a negative slope indicates a negative correlation; a slope of 0 or 1 indicates no correlation. Continuous mineral brittleness index data points are obtained using formula (1), and the elastic parameters are normalized using formula (2). After processing, the dimensions are unified and the correlation does not change. Five elastic parameters, including Young's modulus (E), Poisson's ratio (μ), bulk modulus (K), shear modulus (α), and Lame coefficient (λ), are calculated using formula (3-7). Among them, w QFM is the content of quartz and feldspar in shale, %; w CARM is the content of calcite and dolomite in shale, %; w CM is the clay mineral content in shale, %; w PY is the pyrite content in shale, %; V p is the longitudinal wave velocity; V s is the shear wave velocity; Z i is the standardized value of each elastic parameter, dimensionless; X i is the original value of each elastic parameter; max(X i )、min(X i ) are the maximum and minimum values ​​of each elastic parameter in the target layer. Figure 4 It is a correlation analysis diagram of elastic parameters of the present invention.

[0086]

[0087] For any continuous function within a closed interval, any mapping from n-dimension to m-dimension can be completed by a three-layer BP neural network with a single hidden layer structure. Therefore, the improved BP neural network structure established by the present invention is a three-layer network structure with a single input layer, a single output layer and a single hidden layer. Based on the above elastic parameter sensitivity analysis and comparative study, three elastic parameter data of Young's modulus, Poisson's ratio and shear modulus are constructed as model input layer data, and brittleness index is used as model output layer data. Therefore, it is clear that the number of neurons in the input layer of the network is 3, and the number of neurons in the output layer is 1. Table 1 is the overall error table corresponding to different numbers of hidden layer nodes of the present invention. The number of hidden layer neurons is determined by trial and error. When the number of hidden layer nodes is 20, the error reaches 0.15616. When the number of nodes continues to increase, the error change tends to be stable. Accordingly, the number of neuron nodes in the hidden layer of the model is set to 20. Therefore, the improved BP neural network model structure of the present invention is a three-layer network structure consisting of a single input layer, a single hidden layer, and a single output layer, and the number of neurons in the input layer is set to 3, the number of neurons in the hidden layer is set to 20, and the number of neurons in the output layer is set to 1. Figure 5 It is the error analysis diagram of different hidden layer node numbers of the present invention;

[0088] Please refer to Table 1:

[0089] Table 1 The overall error corresponding to different numbers of hidden layer nodes

[0090]

[0091] During the training process of the BP neural network, the model follows the principle of reducing the error between the target output and the actual output, and continuously updates the weights w and thresholds b of the neurons in each hidden layer and the input layer by back propagating the error of the output layer. Therefore, it is called the "error back propagation algorithm", that is, the BP algorithm (Back Propagation).

[0092] Traditional BP neural networks usually use gradient descent as an optimization algorithm to reduce back propagation errors during model training. The error function (the square of the difference between the model output and the actual value) is minimized by continuously iteratively updating the model parameters to obtain the optimal BP neural network model. The "learning rate" can be represented by the iterative step size in the gradient descent algorithm. It can affect the speed and convergence of the iteration, and directly affect the model training speed and final performance. Error function J(θ) and independent variable parameter θ k The formula is as follows:

[0093] J(θ)=||Y k (θ)-Y'|| 2 2

[0094]

[0095] In the formula, θ is the weight of the input data, Y k (θ) is the predicted output result, Y' is the actual output of the input data, α k is the learning rate when the number of iterations is k, is the gradient of the error function when the number of iterations is k.

[0096] Fixed learning rate has the advantages of simple implementation and easy debugging, and is currently the most widely used learning rate adjustment strategy. Usually, the learning rate is selected in the range of 0.01 to 0.8 by empirical method, or the learning rate is tested in sequence by enumeration method with values ​​of 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, and 1. However, fixed learning rate can easily lead to problems such as long model training time, slow convergence speed in the later stage of model training, or premature convergence of the model and falling into the local optimal solution.

[0097] This study is based on the improved BP neural network shale reservoir brittleness index seismic prediction method, improves the learning rate in the gradient descent method, and uses an adaptive method to determine the learning rate. The improved adaptive learning rate expression is as follows:

[0098]

[0099] In the formula, α is the learning rate, k is the number of iterations, and α 0 is the initial learning rate, a constant, taken as 0.5.

[0100] The adaptive learning rate realizes the function that the learning rate α gradually decreases with the number of training iterations, making the model's parameter optimization and update process smoother during training, avoiding overfitting of the model to the training data, and improving the stability and generalization ability of the model during training.

[0101] The advantage of this improvement is that, on the basis of fully considering the design of different learning rates for different numbers of iterations, the power function is used to ensure that the learning rate generally shows a stable downward trend as the number of iterations increases. Among them, the use of the power function plus 1 makes the initial convergence speed of the model training faster, saving the convergence time of the model when the amount of data is large; using the exponential function The learning rate is controlled to decrease steadily with the increase of the number of iterations and the amount of calculation is small, so as to avoid the learning rate changing too fast and causing model oscillation. Figure 6 It is a comparison diagram of the error convergence curve of the BP neural network of the present invention and the error convergence curve of the improved BP neural network.

[0102] The original data set used in the present invention adopts the actual logging data, rock mechanics experimental data and X-ray diffraction experimental mineral data of the first sub-section of the Longmaxi Formation in the Z block of Yuxi, Sichuan Basin. After data preprocessing, the continuous mineral brittleness index data and the corresponding brittleness index sensitive rock elastic parameter data are obtained. The outliers are eliminated to determine the data set that can be used for model training and verification. The continuous data of 3 wells are randomly selected to participate in the model training. In order to evaluate the prediction performance of the model, the selected model performance evaluation indicators are the mean relative error (MAE) and the determination coefficient (R 2 ). MAE is used to evaluate the difference between the model prediction value and the true value, expressed as a percentage. The smaller the MAE value, the higher the model accuracy. 2 Used to evaluate the correlation between the model prediction value and the true value, R 2 The closer the value is to 1, the more reliable the model is. MAE and R 2 The calculation formula is as follows. The Young's modulus (E), Poisson's ratio (σ), and shear modulus (G) data of the four wells Z1, Z2, Z3, and Z4 in the study area, which have both cross-dipole acoustic logging data and rock and mineral logging interpretation data, are used as the model input characteristic variables, and the corresponding brittleness index data are used as the output characteristic variables. The four wells have a total of 1970 groups of valid data points. Well Z4 is randomly retained as the posterior data, and the remaining 1603 groups of data points of Well Z1, Well Z2, and Well Z3 are used for training and testing of the network system, of which 70% of the data points are used for training and 30% of the data points are used for system automatic testing. The absolute error analysis of the system prediction value and the actual brittleness index shows that the improved BP neural network prediction model has a small relative error, with an average of 3.1042%. Table 2 shows some training data and error analysis data. The correlation analysis between the predicted results of system training data and the actual brittleness data shows that the correlation coefficient between the two is 0.8735, and the training effect is ideal; the correlation analysis between the predicted results of system test data and the actual data shows that the correlation coefficient between the two is 0.8489, and the test effect is good. Figure 7 This is a correlation analysis diagram of the model training, test data and real data of the present invention.

[0103]

[0104] Table 2 Partial training data set and prediction error statistics of the improved BP neural network model

[0105]

[0106] In order to analyze and verify the effect of the improved BP neural network brittleness index prediction model, the data of a well that did not participate in the model training process was used as the model verification data. By analyzing the correlation between the predicted value and the true value of the well, the determination coefficient (R 2) to test and evaluate the prediction effect of the model. The three rock elastic parameters in the 367 data points of the Z4 well that did not participate in the model training and system testing were used as input. The brittleness index output by the neural network was compared with the corresponding actual brittleness index. The predicted value of the vertical brittleness index of the Z4 well in the study area obtained by the improved BP neural network model had a high degree of match with the actual value ( Figure 8 a), the prediction correlation coefficient of 367 data points is 0.811 ( Figure 8 b). Figure 8 This is an analysis diagram for verifying the effect of the improved BP neural network model of the present invention.

[0107] The prestack inversion process includes: prestack seismic data optimization, well seismic fine calibration, and elastic parameter volume calculation. The original CDP gathers were processed by super gathers to obtain high-quality prestack seismic data; then the seismic data was converted into AVA gathers for prestack inversion. The quality of the original CDP gathers and the optimized Super CDP gathers was significantly improved, and the signal-to-noise ratio was significantly improved. Under the constraints of stratigraphic data, well data and geological models, the joint inversion of P-wave velocity, S-wave velocity and density was completed, and the Young's modulus, Poisson's ratio and shear modulus elastic parameter data volume were obtained. In the plane diagrams of each elastic parameter, the distribution of Young's modulus is low in the southwest and high in the northeast; the distribution of Poisson's ratio and shear modulus is high in the southwest and low in the northeast. Moreover, Young's modulus and Poisson's ratio show a negative correlation, which is consistent with the results of the intersection analysis of shale brittleness index. Fig. 9 This is a diagram of pre-stack seismic data optimization and well-seismic fine calibration analysis of the present invention. Fig.10 It is a plane diagram of the pre-stack simultaneous inversion elastic parameter results of the present invention.

[0108] The Young's modulus, Poisson's ratio and shear modulus prediction data of pre-stack inversion are input into the constructed neural network model to obtain the brittleness index prediction result. The plane distribution diagram of the prediction result is shown in Fig.11 As shown in the figure, the average brittleness index is between 60% and 70% according to the improved BP neural network method brittleness index prediction plane result diagram, and the brittleness index plane law shows that it gradually decreases from northwest to southeast. Fig.11 It is a plane distribution diagram of the brittleness index prediction results of the improved BP neural network method of the present invention.

[0109] In order to understand the difference between this method and the traditional method, the prediction results of the conventional Rickman brittleness index prediction method and the prediction results of this method are compared and analyzed. The analysis of the brittleness index prediction plane map of the two methods and the plane interpolation map of the well point brittleness data shows that the actual mineral brittleness index plane distribution map [ Fig.12 (a)] law and the brittleness index prediction results of the improved BP neural network method of the present invention [ Fig.12(b)] is consistent with the above, and the wells have more abundant seismic prediction information. The average brittleness index is predicted to be between 60% and 70%, and the brittleness index plane law shows a trend of gradually decreasing from northwest to southeast. The conventional Rickman method brittleness index prediction plane map [ Fig.12 (c)] There is a certain gap between the law and the plane law of the actual well point brittleness index data. Fig.12 It is a plane comparison diagram of the brittleness index prediction results of the present invention.

[0110] The comparison between the predicted value and the actual value of the average brittleness index of the target layer of a single well and the statistical analysis of the errors show that (Table 3, Fig.13 ), the maximum relative error of the average brittleness index of the well point target layer predicted by the present invention is 3.3783%, while the maximum relative error of the average brittleness index predicted by the Rickman brittleness index prediction method is 8.1924%. Fig.13 It is a comparison chart between the predicted values ​​and measured values ​​of shale brittle earthquake based on the improved BP neural network and the Rickman method of the present invention.

[0111] Table 3 Statistics of relative errors of brittleness index prediction results at well points

[0112]

[0113] The relative errors between the predicted values ​​and the actual values ​​of 1970 data points were counted, and the prediction effects of the two methods were compared and analyzed ( Fig.14 ), if the relative error <±5% is defined as prediction compliance, otherwise it is not consistent, then the improved BP neural network method brittleness index seismic prediction result consistency rate is 95.79%, while the Rickman method brittleness index seismic prediction result consistency rate is 69.16%, the present invention proposes a shale reservoir brittleness index seismic prediction method based on the improved BP neural network, the prediction consistency rate is 26.63% higher than the traditional method. Fig.14 This is a data comparison and analysis chart of the brittleness index prediction results.

[0114] The above is a detailed introduction to the construction of the improved BP neural network model, the inversion of prestack seismic elastic parameters, the seismic prediction of the shale reservoir brittleness index based on the improved BP neural network and the comparative analysis provided in the embodiments of the present invention.

[0115] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A seismic prediction method for shale reservoir brittleness index based on an improved BP neural network, characterized in that: The following steps are involved: S1: Obtain logging data and experimental data of several known wells, wherein the logging data are logging parameters at n different depths in the well, including longitudinal wave velocity Vp, shear wave velocity Vs and density ρ, and the experimental data are various rock parameters, including Young's modulus, Poisson's ratio, shear modulus, Lame coefficient and bulk modulus. The rock dynamic parameters calculated by fitting and correcting the logging data using the experimental data are static rock parameters, providing sufficient data points for neural network model training, and randomly selecting one well as a verification well, and the rest as training wells; S2: Correlation analysis and comparison are performed on each rock parameter in the training well and the mineral brittleness index. The three rock parameters with the top three absolute values ​​of the correlation coefficient are selected as independent variables, marked as independent variable A, independent variable B, and independent variable C respectively, and the mineral brittleness index is used as the dependent variable; S3: Construct the model data set, and use the three independent variables at the same depth of all wells to form sample X i ={X Ai , X Bi , X Ci }, i represents the i-th depth, i = 1, 2, 3, ..., n, X Ai , X Bi , X Ci They represent the values ​​of independent variable A, independent variable B, and independent variable C at the i-th depth respectively. The brittleness index corresponding to this depth is Y i , Y i As the label of the sample, the sample data and the sample label constitute the model data set; S4: Establish a training matrix, and construct all sample data into a matrix X = {X A , X B , X C }, where X A , X B , X C For all X Ai , X Bi , X Ci The column vector composed of the depth corresponding to the brittleness index Y i Construct matrix Y; preset an iteration matrix X k ={X Ak , X Bk , X Ck }, the number of iterations N, k = 1 to N; and when k = 1, X Ak =X A , X Bk =X B , X Ck =X C ; S5: Use trial and error method to calculate the error of different hidden layer node models, determine the number of hidden layer neurons of BP neural network model, and determine the structure of BP neural network model; S6: Improve the learning rate in the gradient descent method and use the improved BP neural network model to obtain the optimal parameters for model training; S7: Select a verification well, obtain its corresponding data in the model data set, input it into the improved BP neural network brittleness index prediction model, output its predicted value and compare it with the true value, and use the determination coefficient evaluation index to verify the model stability; S8: Using the prestack inversion method, the three rock parameter data volumes in S2 are calculated; S9: Based on the improved BP neural network model, the rock parameter data body calculated in S8 is used as the neural network input data to predict the shale reservoir brittleness index data body, and then determine the shale reservoir brittleness index seismic prediction results.

2. The method for seismic prediction of shale reservoir brittleness index based on an improved BP neural network according to claim 1, characterized in that: In S1, the rock parameters Young's modulus (E), Poisson's ratio (μ), bulk modulus (K), shear modulus (G) and Lame coefficient (λ) are calculated using the following formulas: V p is the longitudinal wave velocity, V s is the shear wave velocity.

3. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 1, characterized in that: The S6 at least comprises the following steps: S61: Normalize the elements in the matrix X respectively to obtain the normalized matrix X ’ ; S62: For matrix X k The elements in are normalized respectively to obtain the normalized matrix X k ’ ; S63: Normalize the elements in the matrix Y respectively to obtain the normalized matrix Y ’ ; S64: Calculate the matrix X k ''s predicted value.

4. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 2, characterized in that: The S64 at least includes the following steps: S641: Use Sigmoid activation function to transform the matrix X k 'Input into the hidden layer of the neural network to obtain the hidden layer output matrix X KH '; S642: Output the hidden layer matrix X KH 'Use the Purelin activation function to input to the output layer and get the output result Y k ; S643: Calculate Y k The error between and Y' is denoted as the error function J(θ); S644: Determine the gradient of the error function when the number of iterations is k S645: Using learning rate α k Multiply by the gradient of the error function to get the distance of gradient descent when the current iteration number is k S646: When the number of iterations is k, the model parameter θ k For the updated model training parameters, determine that the value of the error function J(θ) is less than the iteration termination condition ε, then the model training ends, otherwise proceed to step S647; S647: Update learning rate α k After the update is completed, proceed to step S644.

5. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 3 is characterized by: The error function J(θ) and the model parameter θ in S64 k and the learning rate α k The specific calculation is based on the following formula: J(θ)=||Y k (θ)-Y′||2 2 θ is the weight of the input data, Y k (θ) is the predicted output result, Y′ is the actual output of the input data, α k is the learning rate when the number of iterations is k, is the gradient of the error function when the number of iterations is k, α is the learning rate, k is the number of iterations, α0 is the initial learning rate, a constant, which is taken as 0.

5.

6. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 3, characterized in that: In S641-S643, the normalization process is performed using the following formula: Z i is the standardized value of each elastic parameter, dimensionless; X i is the original value of each elastic parameter; max(X i )、min(X i ) are the maximum and minimum values ​​of each elastic parameter in the target layer.

7. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 1, characterized in that: The S7 determination coefficient is calculated using the following formula: n is the number of samples, is the predicted value of the ith sample, yi is the true value of the ith sample, is the average value among the sample prediction values, is the mean of the true values ​​of the samples.

8. The method for seismic prediction of shale reservoir brittleness index based on improved BP neural network according to claim 1, characterized in that: The S8 at least comprises the following steps: S81: Super gather processing is performed on the seismic data, and further converted into pre-stack angle gathers; S82: Extract seismic wavelets and complete the calibration of strong seismic reflection interface and logging curve; S83: Based on the well location information, layer information and seismic wavelet, establish the initial model of P-wave velocity, S-wave velocity and density required for inversion; S84: Use the Zoeppritz equation to perform prestack inversion and obtain the inverted P-wave velocity, S-wave velocity and density data volume; S85: Calculate the rock parameter data using rock mechanics formulas.

Images