Intelligent prediction method for total organic carbon content of shale based on experience constraint

Abstract

The invention discloses a shale total organic carbon content intelligent prediction method based on experience constraint, and relates to the technical field of geophysical logging. After a sample data set comprising a training set and a test set is constructed according to logging information of a research area, a wave impedance curve is reconstructed, multiple pieces of TOC sample data are obtained, a total organic carbon content empirical model is determined based on a DEN-GR correction LogR model and introduced into a fitness function of a GWO algorithm, an SVR model is optimized by utilizing the GWO algorithm and the TOC sample data in the training set, and the total organic carbon content is obtained. And determining an optimal solution of a penalty coefficient and a kernel function parameter in the SVR model, obtaining the trained SVR model, verifying the accuracy of the SVR model for predicting the total organic carbon content of the shale by adopting TOC sample data in the test set, and applying the verified SVR model to intelligent prediction of the total organic carbon content of the shale. According to the method, empirical constraint and the GWO-SVR algorithm are combined, accurate prediction of the total organic carbon content of shale is achieved, and a basis is provided for guiding exploration and development of shale reservoirs.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical logging, and in particular to an intelligent prediction method for total organic carbon content of shale based on empirical constraints. Background Art

[0002] As an important unconventional oil and gas resource, shale oil and gas has become an important strategic successor resource. As an important parameter in shale reservoirs, total organic carbon content (TOC) can directly characterize the generation, storage and production capacity of shale gas / shale oil. Usually, experts collect source rock samples and obtain high-precision TOC content through laboratory analysis and measurement. This method usually requires the sample to be crushed and pyrolysis analysis is performed using instruments such as the Rock-Eval pyrolyzer. Although the TOC content results obtained are direct and accurate, destructive testing of core or cutting samples is required; the other method is to use an elemental analyzer to measure the carbon content in the sample to infer TOC. These two methods are expensive, time-consuming and labor-intensive, and are not conducive to promotion throughout the well section.

[0003] In order to further obtain the distribution of total organic carbon content in continuous vertical strata based on experimental data, experts in this field have proposed a variety of empirical formulas, such as conventional logging curve fitting, △logR method, etc. However, with the increase in oil field reserve and production pressure, the standard of TOC accuracy is also increasing, and there are great limitations in relying solely on empirical evaluation methods. In addition, when faced with complex geological backgrounds and low background contrast conditions, the mapping relationship between conventional logging curves and total organic carbon content is weak when evaluating shale reservoirs.

[0004] The existence of the above problems has limited the application of conventional linear processing and traditional empirical methods in the prediction of shale total organic carbon content. Therefore, how to make full use of existing experimental data and conventional logging curves, combine empirical constraints and the nonlinear processing capabilities of intelligent machine learning, and propose an intelligent prediction method for shale total organic carbon content based on empirical constraints is of great significance for the study of shale total organic carbon content. Summary of the invention

[0005] The present invention aims to solve the above problems and proposes an intelligent prediction method for the total organic carbon content of shale based on empirical constraints. The empirical constraints are combined with the GWO-SVR algorithm. The △LogR model is modified based on DEN-GR and introduced into the GWO-SVR algorithm. By introducing the total organic carbon content empirical model into the model parameter optimization process of the SVR model, the total organic carbon content of shale is accurately predicted, which provides a basis for the exploration and development of shale reservoirs.

[0006] The present invention adopts the following technical solutions:

[0007] The intelligent prediction method of shale total organic carbon content based on experience constraints includes the following steps:

[0008] Step 1: Obtain well logging data in the study area, reconstruct the wave impedance curve, select the well logging curve related to the total organic carbon content, and obtain the total organic carbon content sample data of multiple shales;

[0009] Step 2, constructing a sample data set, and randomly assigning the shale total organic carbon content sample data to the training set and test set of the sample data set;

[0010] Step 3, based on the DEN-GR modified △LogR model, determine the empirical model of total organic carbon content in the study area;

[0011] Step 4: Optimize the SVR model based on the GWO algorithm, introduce the total organic carbon content empirical model into the Gray Wolf optimization process, combine the total organic carbon content sample data of shale in the training set, use the GWO algorithm to determine the optimal model parameters of the SVR model, and obtain the trained SVR model;

[0012] Step 5: Input the shale total organic carbon content sample data in the test set into the trained SVR model, use the trained SVR model to predict the total organic carbon content of the shale total organic carbon content sample data, and verify the accuracy of the trained SVR model.

[0013] Preferably, in step 1, the wave impedance curve is reconstructed according to the acoustic time difference curve and the density curve by using the ratio between the acoustic time difference and the formation density at different depths;

[0014] The logging data of the study area are conventional logging curves of wells logged in the study area, including acoustic time difference curves, density curves, neutron porosity curves, natural gamma curves and resistivity curves. According to the acoustic time difference curves and density curves, the wave impedance curve is reconstructed by using the ratio between the acoustic time difference and the formation density at different depths;

[0015] Based on the total organic carbon content curve of the well, a logging curve that can change with the total organic carbon content is selected from the conventional logging curves, and combined with the wave impedance curve, the total organic carbon content related logging curve is obtained. Based on the total organic carbon content related logging curves and the total organic carbon content curves at different depths, multiple shale total organic carbon content sample data are obtained.

[0016] Preferably, the shale total organic carbon content sample data consists of multiple characteristic parameter values ​​and measured values ​​of the shale total organic carbon content at the same depth, wherein the characteristic parameter values ​​are determined based on the total organic carbon content related logging curves, and the measured values ​​of the shale total organic carbon content are determined based on the total organic carbon content curves.

[0017] Preferably, in step 3, according to the well logging curve of the study area, based on the DEN-GR modified △LogR model, combined with the well logging curve fitting of the study area, the model parameters are determined, and the total organic carbon content empirical model is determined as:

[0018] TOC=[a×lg(GR)+b×DEN+c]×ΔlogR

[0019] in,

[0020] The △LogR model based on DEN-GR correction is:

[0021]

[0022] Where TOC is the total organic carbon content; a, b, c are model parameters; GR is the natural gamma value; DEN is the formation density value; R t is the resistivity, unit is Ω·m; R baseline is the resistivity corresponding to the baseline in the resistivity curve, in Ω·m; Δt is the acoustic time difference, in μs / m; Δt baseline It is the acoustic time difference value corresponding to the baseline in the acoustic time difference curve, in μs / m.

[0023] Preferably, in step 4, optimizing the SVR model based on the GWO algorithm includes the following sub-steps:

[0024] Step 4.1, initialize the GWO algorithm, set the type of gray wolf individuals, wolf pack position and fitness function in the GWO algorithm, and determine the number of search agents, maximum optimization times and initial value of the search space of the GWO algorithm;

[0025] Step 4.2, setting the initial values ​​of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and inputting the shale total organic carbon content sample data in the training set into the SVR model and the total organic carbon content empirical model;

[0026] Step 4.3, based on the GWO algorithm, the optimal search is performed to obtain the solutions found by each individual in the wolf pack, and the mean square error is calculated using the SVR model, and the penalty term is calculated using the total organic carbon content empirical model to determine the fitness function corresponding to each solution;

[0027] Step 4.4, compare the fitness function values ​​of each solution, determine the best solution, and update the position of the wolf pack according to the positions of the Alpha wolf, Beta wolf, and Delta wolf. Re-search based on the GWO algorithm to obtain the solution found by each individual in the wolf pack, and update the coefficient vector and solution when the wolf pack surrounds the prey;

[0028] Step 4.5, repeat steps 4.3 and 4.4 until the preset maximum number of optimization times is reached, stop using the GWO algorithm for optimal search, output the optimal solution of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and obtain the trained SVR model.

[0029] Preferably, the fitness function is:

[0030]

[0031] Wherein, Fitness is fitness; w is input sample; MSE is mean square error; Penalty is the penalty term based on the total organic carbon content empirical model, which is used to ensure the data scale balance between the mean square error and the penalty term; λ is the penalty factor, which is used to weigh the importance of the mean square error MSE and the penalty term; N is the total number of shale total organic carbon content sample data in the training set; i is the sample number; y is the penalty factor ... i is the measured TOC value of the i-th shale total organic carbon content sample data; is the TOC prediction value of the i-th shale total organic carbon content sample data; j is the characteristic parameter number; M is the total number of characteristic parameters; x j is the jth characteristic parameter value; is the TOC prediction value for the characteristic parameter value x j The partial derivative of is used to express the rate of change of the TOC prediction value relative to the jth characteristic parameter; The partial derivative of the TOC empirical value calculated by the total organic carbon content empirical model with respect to the j-th characteristic parameter is used to represent the rate of change of the TOC empirical value relative to the j-th characteristic parameter.

[0032] Preferably, in the GWO algorithm, individuals in the wolf pack are divided into four categories according to social levels, namely Alpha wolf, Beta wolf, Delta wolf and Omega wolf from high to low social levels, wherein the Alpha wolf is the leader of the wolf pack, ranking at the first level of the wolf pack, responsible for decision-making and guiding the actions of the wolf pack; the Beta wolf ranks at the second level of the wolf pack, used to assist the Alpha wolf in making decisions and guide the actions of other low-level wolves; the Delta wolf ranks at the third level of the wolf pack, responsible for reconnaissance and vigilance; the Omega wolf ranks at the fourth level of the wolf pack, used to execute the orders of the Alpha wolf, Beta wolf and Delta wolf;

[0033] When performing optimal search based on the GWO algorithm, the solution searched by the Alpha wolf is the best solution, the solution searched by the Beta wolf is the second best solution, the solution searched by the Delta wolf is the third best solution, and the solution searched by the Omega wolf is the remaining alternative solution;

[0034] The collective hunting process of the wolf pack adopts global search and local search alternately, and includes four stages, namely, surrounding prey, hunting, attacking prey and searching prey;

[0035] The expression for surrounding the prey is:

[0036]

[0037] In the formula, t is the current optimization times; is the distance vector between the gray wolf and the prey; The distance vector between the gray wolf and the prey during the next optimization; are coefficient vectors; is the position vector of the prey; is the position vector of the gray wolf; It indicates that it decreases linearly from 2 to 0 during the optimization process; are all random vectors in [0,1];

[0038] The expression of the hunting is:

[0039]

[0040]

[0041] In the formula They are the distance vectors of Alpha wolf, Beta wolf, and Delta wolf respectively; is the coefficient vector of Alpha wolf, is the coefficient vector of Beta wolf, is the coefficient vector of Delta wolf, Both are used to adjust the distance between individuals in a wolf pack; They are the current positions of the Alpha wolf, Beta wolf, and Delta wolf respectively; The new potential position of Omega wolf is updated according to the position and distance of Alpha wolf. The new potential position of Omega wolf is updated according to the position and distance of Beta wolf. The new potential position of Omega wolf is updated based on the position and distance of Delta wolf; To control the coefficient vector determined by the position change of the Omega wolf according to the Alpha wolf, is the coefficient vector determined by the position change of Omega wolf according to Beta wolf, is the coefficient vector determined by the position change of Omega wolf according to Delta wolf; The final position of the Omega Wolf. Omega wolves tend to stray from their prey when Omega wolves tend to move closer to their prey.

[0042] Preferably, the SVR model selects radial basis function as the kernel function to deal with complex nonlinear problems; the model parameters of the SVR model include a penalty coefficient C and a kernel function parameter gamma, wherein the penalty coefficient C determines the penalty intensity of the SVR model for misclassification and is used to control the tolerance of the SVR model to errors; the kernel function parameter gamma determines the radiation range of the SVR model and is used to control the influence range of a single sample data.

[0043] Preferably, in step 5, the shale total organic carbon content sample data in the test set is input into the trained SVR model, and the trained SVR model is used to predict the shale total organic carbon content based on the shale total organic carbon content sample data to obtain the predicted value, and compare it with the measured value of the shale total organic carbon content in the shale total organic carbon content sample data, calculate the accuracy of the prediction of the trained SVR model and compare it with the preset accuracy value, if the accuracy of the prediction of the trained SVR model is lower than the preset accuracy value, return to step 4 to continue training the SVR model, otherwise, obtain the verified SVR model, and directly use the verified SVR model to predict the shale total organic carbon content.

[0044] The present invention has the following beneficial effects:

[0045] The present invention proposes an intelligent prediction method for shale total organic carbon content based on empirical constraints. The method uses TOC experimental data and conventional logging curves as basic data support, and based on the DEN-GR modified △LogR model, determines the total organic carbon content empirical model of the study area and introduces it into the GWO algorithm. By introducing the empirical penalty term into the fitness function of the GWO algorithm to constrain the consistency of the predicted value with the overall trend of the measured shale total organic carbon content, the GWO algorithm is used to optimize the parameters of the SVR model. The method of the present invention combines empirical constraints with machine learning, overcomes the limitation of the current shale total organic carbon content prediction method when applied to shale reservoir evaluation due to the geological background environment, and effectively improves the accuracy and reliability of shale total organic carbon content prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The present invention is a flow chart of an intelligent prediction method for total organic carbon content in shale based on empirical constraints.

[0047] Figure 2 Flow chart of the GWO algorithm of the present invention.

[0048] Figure 3 This is the predicted diagram of the total organic carbon content of shale in Well Y in Example 2. DETAILED DESCRIPTION

[0049] The following takes a certain research area as an example and further describes the specific implementation of the present invention in combination with the accompanying drawings:

[0050] Example 1

[0051] The present invention proposes an intelligent prediction method for total organic carbon content in shale based on empirical constraints, such as Figure 1 As shown, the specific steps include:

[0052] Step 1: Select a shale reservoir area as the study area, obtain the logging data of the study area, that is, the conventional logging curves of the wells in the study area, including the acoustic time difference curve, density curve, neutron porosity curve, natural gamma curve and resistivity curve. According to the acoustic time difference curve and density curve, the ratio between the acoustic time difference and the formation density at different depths is used to reconstruct the wave impedance curve to reflect the physical properties of the rock.

[0053] Then, based on the logged total organic carbon (TOC) content curve, a logging curve that can change with the total organic carbon content is selected from the conventional logging curves, and combined with the wave impedance curve, the total organic carbon content related logging curve is obtained. Based on the total organic carbon content related logging curves and the total organic carbon content curves at different depths, multiple shale total organic carbon content sample data are obtained.

[0054] The shale total organic carbon content sample data consists of multiple characteristic parameter values ​​at the same depth and the actual measured value of the shale total organic carbon content, wherein the characteristic parameter value is determined according to the total organic carbon content related logging curve, and the actual measured value of the shale total organic carbon content is determined according to the total organic carbon content curve.

[0055] Step 2: construct a sample data set and randomly assign the shale total organic carbon content sample data to the training set and test set of the sample data set, where the training set is used to train the SVR model, and the test set is used to test the accuracy of the shale total organic carbon content prediction of the trained SVR model.

[0056] Step 3: Determine the empirical model of total organic carbon content in the study area based on the DEN-GR modified △LogR model.

[0057] The traditional △LogR model is a TOC calculation method based on acoustic time difference, resistivity and maturity parameters proposed by Passey et al. in 1990 based on Archie's formula. The traditional △LogR model has a large error when applied to the study of TOC content of deep source rocks. Therefore, in this embodiment, the traditional △LogR model is modified based on DEN-GR, and fitted with the well logging data of the study area to determine the total organic carbon content empirical model of the study area:

[0058] TOC=[a×lg(GR)+b×DEN+c]×ΔlogR

[0059] in,

[0060] The △LogR model based on DEN-GR correction is:

[0061]

[0062] Where TOC is the total organic carbon content; a, b, c are model parameters; GR is the natural gamma value; DEN is the formation density value; R t is the resistivity, unit is Ω·m; R baseline is the resistivity corresponding to the baseline in the resistivity curve, in Ω·m; Δt is the acoustic time difference, in μs / m; Δt baseline It is the acoustic time difference value corresponding to the baseline in the acoustic time difference curve, in μs / m.

[0063] This embodiment takes well X in the study area as an example, and determines the resistivity R corresponding to the baseline in the resistivity curve according to the well logging data of the study area. baseline =1Ω·m, the acoustic time difference value Δt corresponding to the baseline in the acoustic time difference curve baseline The total organic carbon content of well X is 80 μs / m. Based on the well logging data in the study area, the empirical model for total organic carbon content of well X is determined as:

[0064] TOC=[9.67×lg(GR)-10.62×DEN+9.19]×ΔlogR.

[0065] Step 4: Optimize the SVR model based on the GWO algorithm, introduce the total organic carbon content empirical model into the Gray Wolf optimization process, combine the shale total organic carbon content sample data in the training set, use the GWO algorithm to determine the optimal model parameters of the SVR model, and obtain the trained SVR model.

[0066] Optimizing the SVR model based on the GWO algorithm includes the following sub-steps:

[0067] Step 4.1, initialize the GWO algorithm (i.e., gray wolf optimization algorithm), set the type of gray wolf individuals, wolf pack position, and fitness function in the GWO algorithm, and determine the number of search agents, the maximum number of optimizations, and the initial value of the search space of the GWO algorithm.

[0068] In this embodiment, the GWO algorithm can automatically adjust the model parameters in the SVR (support vector regression) model according to the characteristics and internal structure of the data set, specifically adjusting the penalty coefficient C and kernel function parameter gamma of the SVR model. The goal of the SVR model is to determine the optimal hyperplane in a given data set so that the distance between all data points in the data set and the optimal hyperplane is as small as possible, wherein the kernel function, the penalty coefficient C and the kernel function parameter gamma are key parameters of the SVR model. In this embodiment, the radial basis function (RBF) is selected as the kernel function to handle complex nonlinear problems; the penalty coefficient C determines the penalty intensity of the SVR model for misclassification, which is used to control the tolerance of the SVR model to errors. The larger the value of the penalty coefficient C, the heavier the penalty of the SVR model for errors; the kernel function parameter gamma determines the radiation range of the SVR model, which is used to control the influence range of a single sample data. If the kernel function parameter gamma is too large, it may lead to overfitting, and if it is too small, it may make the model too simple; the GWO algorithm is used to optimize and determine the penalty coefficient C and the kernel function parameter gamma of the SVR model, and the complexity of the SVR model and the accuracy of the prediction results are balanced by adjusting the penalty coefficient C and the kernel function parameter gamma.

[0069] Furthermore, the total organic carbon content empirical model is used to determine the consistency of the overall trend of the predicted value of the penalty term constrained SVR model and the calculated value of the TOC empirical model based on the empirical constraint △LogR model. The fitness function in this embodiment is set as:

[0070]

[0071] Wherein, Fitness is fitness; w is input sample; MSE is mean square error; Penalty is a penalty term based on the total organic carbon content empirical model, which is used to ensure the data scale balance between the mean square error and the penalty term; λ is a penalty factor, which is used to weigh the importance of the mean square error MSE and the penalty term, and ensure the data scale balance between the mean square error and the penalty term. It is generally determined by cross-validation. In this embodiment, five-fold cross-validation is used to determine that the penalty factor λ is 0.1; N is the total number of shale total organic carbon content sample data in the training set; i is the sample number; y i is the measured TOC value of the i-th shale total organic carbon content sample data; is the TOC prediction value of the i-th shale total organic carbon content sample data; j is the characteristic parameter number; m is the total number of characteristic parameters; x j is the jth characteristic parameter value; is the TOC prediction value for the characteristic parameter value x j The partial derivative of is used to express the rate of change of the TOC prediction value relative to the jth characteristic parameter; The partial derivative of the TOC empirical value calculated by the total organic carbon content empirical model with respect to the j-th characteristic parameter is used to represent the rate of change of the TOC empirical value relative to the j-th characteristic parameter.

[0072] Step 4.2, set the initial values ​​of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and input the shale total organic carbon content sample data in the training set into the SVR model and the total organic carbon content empirical model.

[0073] Step 4.3, based on the GWO algorithm, the optimal search is used to obtain the solutions found by each individual in the wolf pack, and the mean square error is calculated using the SVR model. The penalty term is calculated using the total organic carbon content empirical model to determine the fitness function corresponding to each solution.

[0074] In this embodiment, in the GWO algorithm, individuals in the wolf pack are divided into four categories according to social levels, namely Alpha wolf, Beta wolf, Delta wolf and Omega wolf from high to low social levels, among which Alpha wolf is the leader of the wolf pack, ranking first in the wolf pack, responsible for decision-making and guiding the actions of the wolf pack; Beta wolf ranks second in the wolf pack, assisting Alpha wolf in decision-making and guiding the actions of other low-level wolves; Delta wolf ranks third in the wolf pack, responsible for reconnaissance and vigilance; Omega wolf ranks fourth in the wolf pack, executing the orders of Alpha wolf, Beta wolf and Delta wolf.

[0075] When performing optimal search based on the GWO algorithm, the solution searched by the Alpha wolf is the best solution, the solution searched by the Beta wolf is the second best solution, the solution searched by the Delta wolf is the third best solution, and the solution searched by the Omega wolf is the remaining alternative solution.

[0076] The collective hunting process of the wolf pack adopts global search and local search alternately, and includes four stages, namely, surrounding prey, hunting, attacking prey and searching prey, wherein the expression of surrounding prey is:

[0077]

[0078] In the formula, t is the current optimization times; is the distance vector between the gray wolf and the prey; The distance vector between the gray wolf and the prey during the next optimization; are coefficient vectors; is the position vector of the prey; is the position vector of the gray wolf; It indicates that it decreases linearly from 2 to 0 during the optimization process; are all random vectors in [0,1].

[0079] The expression of the hunting is:

[0080]

[0081]

[0082] In the formula, They are the distance vectors of Alpha wolf, Beta wolf, and Delta wolf respectively; is the coefficient vector of Alpha wolf, is the coefficient vector of Beta wolf, is the coefficient vector of Delta wolf, Both are used to adjust the distance between individuals in a wolf pack; They are the current positions of the Alpha wolf, Beta wolf, and Delta wolf respectively; The new potential position of Omega wolf is updated according to the position and distance of Alpha wolf. The new potential position of Omega wolf is updated according to the position and distance of Beta wolf. The new potential position of Omega wolf is updated based on the position and distance of Delta wolf; To control the coefficient vector determined by the position change of the Omega wolf according to the Alpha wolf, is the coefficient vector determined by the position change of Omega wolf according to Beta wolf, is the coefficient vector determined by the position change of Omega wolf according to Delta wolf; For the final position of Omega Wolf, Omega wolves tend to stray from their prey when Omega wolves tend to move closer to their prey.

[0083] Step 4.4, compare the fitness function values ​​of each solution, determine the best solution, and update the position of the wolf pack according to the positions of the Alpha wolf, Beta wolf, and Delta wolf. Re-search based on the GWO algorithm to obtain the solution found by each individual in the wolf pack, and update the coefficient vector and solution when the wolf pack surrounds the prey.

[0084] Step 4.5, repeat steps 4.3 and 4.4 until the preset maximum number of optimization times is reached, stop using the GWO algorithm for optimal search, output the optimal solution of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and obtain the trained SVR model.

[0085] In this embodiment, each time the optimization search is performed based on the GWO algorithm, Figure 2 As shown in the figure, a wolf pack search is first used to obtain multiple solutions. After the best solution is obtained by evaluating the fitness function of each solution and sorting them, the wolf pack position is reset according to the positions of the Alpha wolf, Beta wolf, and Delta wolf, and the coefficient vector and solution are updated and evaluated until the preset maximum number of optimizations is reached, that is, the penalty coefficient C and kernel function parameter gamma in the best SVR model are determined, and the SVR model is set according to the optimal solution of the penalty coefficient C and kernel function parameter gamma.

[0086] Step 5: Input the shale total organic carbon content sample data in the test set into the trained SVR model, use the trained SVR model to predict the total organic carbon content of the shale total organic carbon content sample data, and verify the accuracy of the trained SVR model.

[0087] In this embodiment, the shale total organic carbon content sample data in the test set is input into the trained SVR model, and the trained SVR model is used to predict the shale total organic carbon content based on the shale total organic carbon content sample data to obtain the predicted value of the shale total organic carbon content, and compared with the actual measured value of the shale total organic carbon content in the shale total organic carbon content sample data, and the accuracy of the prediction of the trained SVR model is calculated and compared with the preset accuracy value. If the accuracy of the prediction of the trained SVR model is lower than the preset accuracy value, return to step 4 to continue training the SVR model. Otherwise, a verified SVR model is obtained, and the verified SVR model is directly used to predict the shale total organic carbon content.

[0088] Example 2

[0089] In this embodiment, the traditional empirical method based on the △LogR model, the Grey Wolf Optimization Algorithm-Support Vector Machine Algorithm GWO-SVR and the intelligent prediction method for shale total organic carbon content based on empirical constraints proposed in Example 1 are used to predict the shale total organic carbon content of Well Y, and the prediction results of the shale total organic carbon content of Well Y are obtained, as shown in FIG. Figure 3 shown.

[0090] Figure 3 In the figure, the TOClogR curve is the total organic carbon content curve of shale predicted by the traditional empirical method, the GWO-SVR curve is the total organic carbon content curve of shale predicted by the Grey Wolf Optimization Algorithm-Support Vector Machine Algorithm, and the EC-GWO-SVR curve is the total organic carbon content curve of shale predicted by the intelligent prediction method for total organic carbon content of shale based on empirical constraints proposed in the present invention.

[0091] Comparative analysis shows that the method of the present invention uses the Grey Wolf algorithm based on empirical constraints to optimize the support vector regression model. By introducing the △LogR model based on empirical constraints into the Grey Wolf algorithm to optimize the support vector regression model, based on the existing experimental data and conventional logging curves in the study area, the empirical constraints are combined with machine learning, and the total organic carbon content of shale in the entire well section to be logged is accurately predicted, which is beneficial to guiding the exploration and development of shale reservoirs.

[0092] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. An intelligent prediction method for total organic carbon content in shale based on empirical constraints, characterized in that: The following steps are involved: Step 1: Obtain well logging data in the study area, reconstruct the wave impedance curve, select the well logging curve related to the total organic carbon content, and obtain the total organic carbon content sample data of multiple shales; Step 2, constructing a sample data set, and randomly assigning the shale total organic carbon content sample data to the training set and test set of the sample data set; Step 3, based on the DEN-GR modified △LogR model, determine the empirical model of total organic carbon content in the study area; Step 4: Optimize the SVR model based on the GWO algorithm, introduce the total organic carbon content empirical model into the Gray Wolf optimization process, combine the total organic carbon content sample data of shale in the training set, use the GWO algorithm to determine the optimal model parameters of the SVR model, and obtain the trained SVR model; Step 5: Input the shale total organic carbon content sample data in the test set into the trained SVR model, use the trained SVR model to predict the total organic carbon content of the shale total organic carbon content sample data, and verify the accuracy of the trained SVR model.

2. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 1 is characterized in that: In the step 1, based on the acoustic time difference curve and the density curve, the wave impedance curve is reconstructed by using the ratio between the acoustic time difference and the formation density at different depths; The logging data of the study area are conventional logging curves of wells logged in the study area, including acoustic time difference curves, density curves, neutron porosity curves, natural gamma curves and resistivity curves. According to the acoustic time difference curves and density curves, the wave impedance curve is reconstructed by using the ratio between the acoustic time difference and the formation density at different depths; Based on the total organic carbon content curve of the well, a logging curve that can change with the total organic carbon content is selected from the conventional logging curves, and combined with the wave impedance curve, the total organic carbon content related logging curve is obtained. Based on the total organic carbon content related logging curves and the total organic carbon content curves at different depths, multiple shale total organic carbon content sample data are obtained.

3. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 2 is characterized in that: The shale total organic carbon content sample data consists of multiple characteristic parameter values ​​at the same depth and the actual measured value of the shale total organic carbon content, wherein the characteristic parameter value is determined according to the total organic carbon content related logging curve, and the actual measured value of the shale total organic carbon content is determined according to the total organic carbon content curve.

4. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 1 is characterized in that: In step 3, according to the well logging curve of the study area, based on the DEN-GR modified △LogR model, combined with the well logging curve fitting of the study area, the model parameters are determined, and the total organic carbon content empirical model is determined as follows: TOC=[a×lg(GR)+b×DEN+c]×ΔlogR in, The △LogR model based on DEN-GR correction is: Where TOC is the total organic carbon content; a, b, c are model parameters; GR is the natural gamma value; DEN is the formation density value; R t is the resistivity, unit is Ω·m; R baseline is the resistivity corresponding to the baseline in the resistivity curve, in Ω·m; Δt is the acoustic time difference, in μs / m; Δt baseline It is the acoustic time difference value corresponding to the baseline in the acoustic time difference curve, in μs / m.

5. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 4 is characterized in that: In step 4, optimizing the SVR model based on the GWO algorithm includes the following sub-steps: Step 4.1, initialize the GWO algorithm, set the type of gray wolf individuals, wolf pack position and fitness function in the GWO algorithm, and determine the number of search agents, maximum optimization times and initial value of the search space of the GWO algorithm; Step 4.2, setting the initial values ​​of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and inputting the shale total organic carbon content sample data in the training set into the SVR model and the total organic carbon content empirical model; Step 4.3, based on the GWO algorithm, the optimal search is performed to obtain the solutions found by each individual in the wolf pack, and the mean square error is calculated using the SVR model, and the penalty term is calculated using the total organic carbon content empirical model to determine the fitness function corresponding to each solution; Step 4.4, compare the fitness function values ​​of each solution, determine the best solution, and update the position of the wolf pack according to the positions of the Alpha wolf, Beta wolf, and Delta wolf. Re-search based on the GWO algorithm to obtain the solution found by each individual in the wolf pack, and update the coefficient vector and solution when the wolf pack surrounds the prey; Step 4.5, repeat steps 4.3 and 4.4 until the preset maximum number of optimization times is reached, stop using the GWO algorithm for optimal search, output the optimal solution of the penalty coefficient C and the kernel function parameter gamma in the SVR model, and obtain the trained SVR model.

6. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 5 is characterized in that: The fitness function is: Wherein, Fitness is fitness; w is input sample; MSE is mean square error; Penalty is the penalty term based on the total organic carbon content empirical model, which is used to ensure the data scale balance between the mean square error and the penalty term; λ is the penalty factor, which is used to weigh the importance of the mean square error (MSE) and the penalty term; N is the total number of shale total organic carbon content sample data in the training set; i is the sample number; y i is the measured TOC value of the i-th shale total organic carbon content sample data; is the TOC prediction value of the i-th shale total organic carbon content sample data; j is the characteristic parameter number; M is the total number of characteristic parameters; x j is the jth characteristic parameter value; is the TOC prediction value for the characteristic parameter value x j The partial derivative of is used to express the rate of change of the TOC prediction value relative to the jth characteristic parameter; The partial derivative of the TOC empirical value calculated by the total organic carbon content empirical model with respect to the j-th characteristic parameter is used to represent the rate of change of the TOC empirical value relative to the j-th characteristic parameter.

7. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 5 is characterized in that: In the GWO algorithm, individuals in a wolf pack are divided into four categories according to social levels, namely Alpha wolf, Beta wolf, Delta wolf and Omega wolf from high to low social levels. Among them, Alpha wolf is the leader of the wolf pack, ranking first in the wolf pack, responsible for decision-making and guiding the actions of the wolf pack; Beta wolf ranks second in the wolf pack, assisting Alpha wolf in making decisions and guiding the actions of other low-level wolves; Delta wolf ranks third in the wolf pack, responsible for reconnaissance and vigilance; Omega wolf ranks fourth in the wolf pack, responsible for executing the orders of Alpha wolf, Beta wolf and Delta wolf; When performing optimal search based on the GWO algorithm, the solution searched by the Alpha wolf is the best solution, the solution searched by the Beta wolf is the second best solution, the solution searched by the Delta wolf is the third best solution, and the solution searched by the Omega wolf is the remaining alternative solution; The collective hunting process of the wolf pack adopts global search and local search alternately, and includes four stages, namely, surrounding prey, hunting, attacking prey and searching prey; The expression for surrounding the prey is: In the formula, t is the current optimization times; is the distance vector between the gray wolf and the prey; The distance vector between the gray wolf and the prey during the next optimization; are coefficient vectors; is the position vector of the prey; is the position vector of the gray wolf; It indicates that it decreases linearly from 2 to 0 during the optimization process; are all random vectors in [0,1]; The expression of the hunting is: In the formula, They are the distance vectors of Alpha wolf, Beta wolf, and Delta wolf respectively; is the coefficient vector of Alpha wolf, is the coefficient vector of Beta wolf, is the coefficient vector of Delta wolf, Both are used to adjust the distance between individuals in a wolf pack; They are the current positions of the Alpha wolf, Beta wolf, and Delta wolf respectively; The new potential position of Omega wolf is updated according to the position and distance of Alpha wolf. The new potential position of Omega wolf is updated according to the position and distance of Beta wolf. The new potential position of Omega wolf is updated based on the position and distance of Delta wolf; To control the coefficient vector determined by the position change of the Omega wolf according to the Alpha wolf, is the coefficient vector determined by the position change of Omega wolf according to Beta wolf, is the coefficient vector determined by the position change of Omega wolf according to Delta wolf; For the final position of Omega Wolf, Omega wolves tend to stray from their prey when Omega wolves tend to move closer to their prey.

8. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 5 is characterized in that: The SVR model selects radial basis function as kernel function to deal with complex nonlinear problems; the model parameters of the SVR model include penalty coefficient C and kernel function parameter gamma, wherein the penalty coefficient C determines the penalty intensity of the SVR model for misclassification and is used to control the tolerance of the SVR model to errors; the kernel function parameter gamma determines the radiation range of the SVR model and is used to control the influence range of a single sample data.

9. The intelligent prediction method for total organic carbon content in shale based on empirical constraints according to claim 1 is characterized in that: In step 5, the shale total organic carbon content sample data in the test set is input into the trained SVR model, and the trained SVR model is used to predict the predicted value of the shale total organic carbon content based on the shale total organic carbon content sample data, and compared with the actual measured value of the shale total organic carbon content in the shale total organic carbon content sample data, and the accuracy of the prediction of the trained SVR model is calculated and compared with the preset accuracy value. If the accuracy of the prediction of the trained SVR model is lower than the preset accuracy value, return to step 4 to continue training the SVR model. Otherwise, a verified SVR model is obtained, and the verified SVR model is directly used to predict the shale total organic carbon content.

Images