An artificial intelligence prediction method for shale gas content based on machine learning

Abstract

The present invention discloses an artificial intelligence prediction method for shale gas content based on machine learning, comprising the following steps: step 1, eliminating abnormal values ​​of the measured value of core gas content, and normalizing the P-wave velocity, density, natural gamma and gas content measured values ​​respectively; step 2, introducing relaxation variables, and using data mapping to build a support vector regression prediction model; step 3, using the P-wave velocity and density as input, the shale core gas content as output, using the support vector regression prediction model, and obtaining the gas content prediction value according to leave-one-out cross validation; or using the natural gamma, P-wave velocity, and density as input, the shale core gas content as output, using the support vector regression prediction model, and obtaining the gas content prediction value according to leave-one-out cross validation. The present invention has high accuracy in calculating shale gas content for core, well logging and seismic data, and has high generalization ability and reliability.

Description

Technical Field

[0001] The present invention belongs to the field of earth science technology, is suitable for calculating shale gas content based on core, well logging and seismic data, can provide support for shale gas exploration and development, and particularly relates to an artificial intelligence prediction method for shale gas content based on machine learning. Background Art

[0002] Shale gas is distributed over a large and continuous area, with a large resource scale, but the single well production is low, the production cycle is long, and the recovery efficiency cannot be guaranteed. Accurately predicting the spatial variation of shale reservoir geological information parameters can provide a reliable basis for natural gas extraction and improve the natural gas extraction rate of shale reservoirs.

[0003] At present, the accuracy of calculating gas content based on logging and seismic data is insufficient, which is not conducive to reservoir evaluation and shale gas exploration and development. The basic basis for calculating shale gas content using logging or seismic data mainly comes from statistical analysis, regression fitting or empirical formulas of core test parameters. In short, gas content inversion or other direct and indirect calculation methods are also guided by core test data. Therefore, the key to improving the calculation of gas content based on logging or seismic data is the accurate testing and precise analysis of core gas content. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a shale gas content artificial intelligence prediction method based on machine learning, which can improve the prediction accuracy and reduce the prediction time by applying machine learning to predict shale reservoir parameters.

[0005] The object of the present invention is achieved through the following technical solution: a shale gas content artificial intelligence prediction method based on machine learning, comprising the following steps:

[0006] Step 1: Eliminate the abnormal values ​​of the measured values ​​of the core gas content, and normalize the measured values ​​of the P-wave velocity, the S-wave velocity, the density, the natural gamma and the gas content respectively;

[0007] Step 2: Introduce slack variables and use data mapping to build a support vector regression prediction model;

[0008] Step 3: Taking the P-wave velocity and S-wave density as input and the gas content of the shale core as output, the support vector regression prediction model is used to obtain the predicted value of the gas content according to the leave-one-out cross validation;

[0009] Alternatively, natural gamma, P-wave velocity, and density are used as inputs, and the gas content of shale core is used as output. The support vector regression prediction model is used to obtain the predicted value of gas content based on leave-one-out cross validation.

[0010] Furthermore, the detailed process of step 2 is as follows: define the loss function of the predicted value and the true value of the gas content, do not calculate the loss within the range of the relaxation variable, and only calculate the loss when the error is greater than the relaxation variable; use the kernel function of support vector regression to map the non-linearly separable data to a high-dimensional space, and build a support vector regression prediction model.

[0011] Furthermore, the specific implementation method of step 3 is: suppose all the data sets this time are D = {(x1, y1), (x2, y2), ..., (x m ,y m )}, where x i is the vector of normalized longitudinal wave velocity, shear wave velocity and density, or the vector of normalized natural gamma, longitudinal wave velocity and density; y i is the gas content of the core, m is the total number of samples;

[0012] In support vector regression, the nonlinear mapping function φ(x) is used to map the original sample into a higher-dimensional feature space, thereby achieving the purpose of linear separability; the model corresponding to the partition hyperplane in this feature space is expressed as:

[0013] f(x)=w T φ(x)+b (1)

[0014] Where f(x) is the predicted gas content; w and b are model parameters, the former is the weight and the latter is the intercept;

[0015] Introducing the hard interval ε, according to the principle of minimizing structural risk, the process of solving f(x) is equivalent to solving:

[0016]

[0017] In the formula, the first half is the regularization term, and the second half is the loss function; C is the penalty factor, which is used to control the sample fitting accuracy. The larger its value, the more attention is paid to outliers; l ε is an insensitive loss function with respect to the hard interval ε, and its specific expression is as follows:

[0018]

[0019] Introducing the slack variable ξ i and Rewrite (2) as:

[0020]

[0021] The above equation must be solved under the following constraints:

[0022]

[0023] Introducing the Lagrange multiplier μ i , α i , According to the Lagrange multiplier method, the Lagrange function is:

[0024]

[0025] According to the Lagrange algorithm, the original form of the support vector regression objective function is:

[0026]

[0027] According to the Lagrange dual algorithm, the problem is transformed into an equivalent dual problem:

[0028]

[0029] Find w, b, ξ i , Optimize the minimum value of the function under the conditions, and then find the Lagrange multiplier μ i , α i , The maximum value under ; that is, the Lagrangian function For w, b, ξ i , Find the partial derivative and set it to 0:

[0030]

[0031] Substitute the above formula into (8) and according to the following KKT conditions:

[0032]

[0033] The nonlinear mapping SVR expression is:

[0034]

[0035] In the formula, k(x i ,x)=φ(x i ) T φ(x) is the kernel function, and the radial basis function, i.e., RBF kernel function, is selected as shown in the following formula:

[0036] k(x i ,x)=exp(-||x i -x|| 2 / 2σ 2 ) (12).

[0037] The beneficial effects of the present invention are: due to the high cost of obtaining cores, the small amount of core data obtained, and the lack of a clear relationship between the parameters obtained from logging and the gas content of the cores, the current prediction accuracy of the gas content of the cores based on the empirical formula method is low. At present, machine learning algorithms are becoming more mature, and applying machine learning to predict shale reservoir parameters can improve prediction accuracy and reduce prediction time. The present invention has a high accuracy in calculating the shale gas content based on core, logging and seismic data, has high generalization ability and reliability, and can provide methodological theory and technical support for site selection and layer selection in shale gas exploration areas, drilling deployment, horizontal fracturing stage optimization, and reserve and production construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the prediction method of the present invention;

[0039] Figure 2 It is a cross-plot of P- and S-wave velocities, density and core gas content.

[0040] Figure 3 It is a cross-plot of natural gamma, P-wave velocity, density and gas content.

[0041] Figure 4 It is a gas content curve predicted by various algorithms based on P-wave and S-wave velocity and density.

[0042] Figure 5 It is a bar chart of the determination coefficient and mean square error of each algorithm based on longitudinal and transverse wave velocity and density.

[0043] Figure 6 It is a bar chart of the determination coefficient and mean square error of each algorithm based on natural gamma, longitudinal wave velocity, and density. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0045] like Figure 1 As shown, an artificial intelligence prediction method for shale gas content based on machine learning of the present invention comprises the following steps:

[0046] Step 1: Data preprocessing: In the process of obtaining the gas content of the core, problems such as improper operation may occur, resulting in abnormal measured values ​​of the gas content of the core. Draw the intersection diagram of the P-wave velocity, density and the gas content of the core, such as Figure 2 As shown; the cross-plot of natural gamma, P-wave velocity, density and core gas content, as shown Figure 3As shown. Because there is a strong linear relationship hidden between density and gas content, in order to reduce the impact of outliers on the prediction results, the strong linear relationship between density and gas content can be used to constrain the measured gas content and eliminate outliers to prevent the introduction of large errors and reduce the impact of outliers. While eliminating outliers with large errors, it is necessary to retain the vast majority of core data (90%). In this embodiment, the measured value of the core gas content deviates from the predicted value by ±1.6 errors as an outlier, and the outliers of the measured value of the core gas content are eliminated; then, in order to reduce the impact of the dimensionless prediction effect, speed up the model convergence speed, improve the model training efficiency, and solve the problems of activation function range limitations in neural networks, the P-wave velocity, density, natural gamma and gas content measured values ​​are normalized respectively:

[0047]

[0048] x' and x are the input data (P- and S-wave velocities, density, natural gamma and measured gas content) before and after normalization, respectively. max(x) and min(x) are the maximum and minimum values ​​of the input data, respectively.

[0049] Step 2: Introduce slack variables and use data mapping to build a support vector regression prediction model; the detailed process is as follows: define a loss function that does not calculate the loss between the predicted value and the true value within the range of the slack variable, map the non-linearly separable data to a high-dimensional space, and build a support vector regression prediction model;

[0050] Step 3: Taking the P-wave velocity and S-wave density as input and the gas content of the shale core as output, the support vector regression prediction model is used to obtain the predicted value of the gas content according to the leave-one-out cross validation;

[0051] Alternatively, natural gamma, P-wave velocity, and density are used as inputs, and the gas content of shale core is used as output. The support vector regression prediction model is used to obtain the predicted value of gas content based on leave-one-out cross validation.

[0052] The specific implementation method is: the filtered and normalized longitudinal and transverse wave velocities V P ,V S , density RHOB as support vector regression input, or the normalized natural gamma, longitudinal wave velocity V P , density RHOB as input, core gas content as output, introduce Lagrange multipliers, traverse and differentiate the four parameters of the Lagrange function, and transform it into a dual problem to find the KKT point to obtain the predicted value; the detailed derivation process is as follows:

[0053] Assume that all data sets are D = {(x1, y1), (x2, y2), ..., (x m ,y m )}, where x iis the vector of normalized longitudinal wave velocity, shear wave velocity and density, or the vector of normalized natural gamma, longitudinal wave velocity and density; y i is the gas content of the core, m is the total number of samples;

[0054] In support vector regression, the nonlinear mapping function φ(x) is used to map the original sample into a higher-dimensional feature space, thereby achieving the purpose of linear separability; the model corresponding to the partition hyperplane in this feature space is expressed as:

[0055] f(x)=w T φ(x)+b (1)

[0056] Where f(x) is the predicted gas content; w and b are model parameters, the former is the weight and the latter is the intercept;

[0057] Introducing the hard interval ε, according to the principle of minimizing structural risk, the process of solving f(x) is equivalent to solving:

[0058]

[0059] In the formula, the first half is the regularization term, and the second half is the loss function; C is the penalty factor, which is used to control the sample fitting accuracy. The larger its value, the more attention is paid to outliers; l ε is an insensitive loss function with respect to the hard interval ε, and its specific expression is as follows:

[0060]

[0061] Introducing the slack variable ξ i and Rewrite (2) as:

[0062]

[0063] The above equation must be solved under the following constraints:

[0064]

[0065] Introducing the Lagrange multiplier μ i , α i , According to the Lagrange multiplier method, the Lagrange function is:

[0066]

[0067] According to the Lagrange algorithm, the original form of the support vector regression objective function is:

[0068]

[0069] According to the Lagrange dual algorithm, the problem is transformed into an equivalent dual problem:

[0070]

[0071] Find w, b, ξ i , Optimize the minimum value of the function under the conditions, and then find the Lagrange multiplier μ i , α i , The maximum value under ; that is, the Lagrangian function For w, b, ξ i , Find the partial derivative and set it to 0:

[0072]

[0073] Substitute the above formula into (8) and according to the following KKT conditions:

[0074]

[0075] The nonlinear mapping SVR expression is:

[0076]

[0077] In the formula, k(x i ,x)=φ(x i ) T φ(x) is the kernel function, and the radial basis function, i.e., RBF kernel function, is selected as shown in the following formula:

[0078] k(x i ,x)=exp(-||x i -x|| 2 / 2σ 2 ) (12).

[0079] The technical effect of the gas content prediction method of the present invention is further verified by experiments below.

[0080] The support vector regression prediction model is compared with regression tree (RT), random forest (RF), BP neural network, convolutional neural network (CNN) and linear regression methods. The experimental process is as follows:

[0081] 1. Establish the experimental model according to the following methods:

[0082] (1) A support vector regression prediction model is established according to the method of the present invention.

[0083] (2) Based on the tree algorithm, internal nodes and leaf nodes are introduced to represent the partition attributes and prediction values ​​respectively. The least squares method is used for partitioning. Directed edges are used to connect different partition units between layers to build a decision regression tree.

[0084] (3) By training multiple decision tree weak models to form a strong model, an integrated algorithm of decision trees is formed. That is, a new training set is randomly extracted from the screened data training set using Booster as the input of multiple decision trees, and finally a random forest regression model is formed.

[0085] (4) Relying on the forward propagation of signals and using gradient update to realize back propagation, the number of hidden layers and neurons is set to build a BP neural network model.

[0086] (5) Relying on the built-in functions of tensorflow, a convolutional neural network model is built using convolutional layers, pooling layers, flattening layers, and fully connected layers.

[0087] 2. Taking the P-wave velocity and S-wave density as input and the gas content of shale core as output, the prediction values ​​of different algorithms are obtained by using methods such as support vector regression (SVR), regression tree (RT), random forest (RF), BP neural network, convolutional neural network (CNN) and linear regression according to leave-one-out cross validation; the specific implementation method is as follows:

[0088] (1) According to the method of step S41 of the present invention, the predicted value of gas content using the support vector regression (SVR) algorithm is obtained.

[0089] (2) The screened and normalized P-wave and S-wave velocities and densities are used as the input of the decision regression tree, and the core gas content is used as the output. According to the following formula, the least squares method is continuously used at the cut points of P-wave and S-wave velocities and density to divide the feature space into different units to form a decision regression tree, and the verification data is input to obtain the predicted gas content:

[0090]

[0091] In the formula, x (j) is the jth characteristic variable; s is the x that minimizes the sum of square errors of the two partitioned regions (j) R1 and R2 are the division areas that minimize the above equation, c1 and c2 are the average values ​​of the prediction parameters of the two areas, respectively; y i is the predicted parameter value. The relationship among R1, R2, c1 and c2 is as follows:

[0092]

[0093] (3) The filtered and normalized P-wave and S-wave velocities and densities are used as random forest inputs, and the core gas content is used as output. After multiple tests, the best max_depth, min_samples_leaf, and n_estimators are found to obtain the predicted gas content based on random forest.

[0094] (4) The filtered and normalized P-wave and S-wave velocities and densities are used as the input of the BP neural network, and the core gas content is used as the output. The neuron connection weights and biases are continuously corrected according to the Adam algorithm and MSE to obtain the predicted value of gas content.

[0095] (5) The filtered and normalized P-wave and S-wave velocities and densities are used as the input of the convolutional neural network, and the core gas content is used as the output. Multiple training samples of size 1×3 are passed to the convolutional layer. The convolutional layer uses the same patch method and uses 128 convolution kernels of size 1×2 to convolve with the training samples to obtain 128 1×3 feature maps, which are input to 128 convolutional layers of size 128×1×2. The same patch method is used to obtain 128 1×3 feature maps, which are input to the pooling layer of size 1×2 to obtain 128 1×2 feature maps, and then input to two convolutional layers containing 256 convolution kernels and one pooling layer to obtain the output and pass it to the next Flatten layer for flattening and then to the fully connected layer to obtain the output. The predicted value and the true value are calculated according to the minimum mean square error loss, which is fed back to the hidden layer neurons to correct the parameters and obtain the gas content prediction value based on the convolutional neural network.

[0096] The mean square error (MSE) and determination coefficient (R) of each model for predicting gas content using input P-wave velocity and S-wave velocity and density are calculated by (9) and (10). 2 ), such as columnar Figure 5 .

[0097]

[0098]

[0099] In formula (9) and (10), m is the total number of samples, is the predicted value of gas content; i is the core test value of gas content, which is also the true value; is the mean of the true value of gas content.

[0100] The above methods are used to predict seismic data, and the actual and predicted gas content values ​​are as follows: Figure 4 shown.

[0101] This embodiment uses the mean square error (MSE) and determination coefficient (R 2 )like Figure 5 shown.

[0102] 3. Take natural gamma, P-wave velocity, and density as input, and shale core gas content as output, and then obtain the predicted gas content using the method in 2. The mean square error (MSE) and determination coefficient (R 2 )like Figure 6 shown.

[0103] from Figure 5 and Figure 6 It can be seen that the mean square error (MSE) of the present invention is smaller than that of the other methods, and the coefficient of determination (R 2 ) are greater than the other methods, proving that the model of the present invention has higher prediction accuracy.

[0104] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A shale gas content artificial intelligence prediction method based on machine learning, characterized in that: The following steps are involved: Step 1: Eliminate the abnormal values ​​of the measured values ​​of the core gas content, and normalize the measured values ​​of the P-wave velocity, the S-wave velocity, the density, the natural gamma and the gas content respectively; Step 2: Introduce relaxation variables and use data mapping to build a support vector regression prediction model; the detailed process is as follows: define the loss function of the predicted value and the true value of the gas content, and do not calculate the loss when the error is less than the sum of the hard interval and the relaxation variable, and only calculate the loss when the error is greater than the sum of the two; use the kernel function of support vector regression to map the non-linearly separable data to a high-dimensional space and build a support vector regression prediction model; Step 3: Taking the P-wave velocity and S-wave density as input and the gas content of the shale core as output, the support vector regression prediction model is used to obtain the predicted value of the gas content according to the leave-one-out cross validation; Alternatively, natural gamma, P-wave velocity, and density are used as inputs, and shale core gas content is used as output, and a support vector regression prediction model is used to obtain a predicted value of gas content based on leave-one-out cross validation; The specific implementation method is: suppose all the data sets this time are D = {(x1, y1), (x2, y2), ..., (x m ,y m )}, where x i is the vector of normalized longitudinal wave velocity, shear wave velocity and density, or the vector of normalized natural gamma, longitudinal wave velocity and density; y i is the gas content of the core, m is the total number of samples; In support vector regression, the nonlinear mapping function φ(x) is used to map the original sample into a higher-dimensional feature space, thereby achieving linear separability. The model corresponding to the division of the hyperplane in this feature space is expressed as: f(x)=w T φ(x)+b (1) Where f(x) is the predicted gas content; w and b are model parameters, the former is the weight and the latter is the intercept; Introducing the hard interval ε, according to the principle of minimizing structural risk, the process of solving f(x) is equivalent to solving: In the formula, the first half is the regularization term, and the second half is the loss function; C is the penalty factor, which is used to control the sample fitting accuracy. The larger its value, the more attention is paid to outliers; l ε is an insensitive loss function with respect to the hard interval ε, and its specific expression is as follows: Introducing the slack variable ξ i and Rewrite (2) as: The above equation must be solved under the following constraints: Introducing the Lagrange multiplier μ i , α i , According to the Lagrange multiplier method, the Lagrange function is: According to the Lagrange algorithm, the original form of the support vector regression objective function is: According to the Lagrange dual algorithm, the problem is transformed into an equivalent dual problem: Find w, b, ξ i , Optimize the minimum value of the function under the conditions, and then find the Lagrange multiplier μ i , α i , The maximum value under ; that is, the Lagrangian function For w, b, ξ i , Find the partial derivative and set it to 0: Substitute the above formula into (8) and according to the following KKT conditions: The nonlinear mapping SVR expression is: In the formula, k(x i ,x)=φ(x i ) T φ(x) is the kernel function, and the radial basis function, i.e., RBF kernel function, is selected as shown in the following formula: k(x i ,x)=exp(-||x i -x|| 2 / 2σ 2 ) (12)。

Images