Shale TOC prediction method based on logging data feature derivation and machine learning

Abstract

The invention discloses a shale TOC prediction method based on logging data feature derivation and machine learning, and relates to the technical field of shale TOC prediction, and the shale TOC prediction method based on logging data feature derivation and machine learning mainly comprises the steps: obtaining shale TOC data through a rock pyrolysis experiment; according to the shale TOC data, using a geological research tool to obtain original logging data and preprocessing the original logging data; and performing feature derivation on the preprocessed logging data to obtain a new data set, training a random forest machine learning model to obtain a shale TOC prediction model, and predicting shale data to be measured to obtain a shale TOC prediction result. By implementing the shale TOC prediction method based on logging data feature derivation and machine learning provided by the invention, the flexibility, applicability and accuracy of shale TOC prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of shale total organic carbon prediction, and more specifically, to a shale TOC prediction method based on well logging data feature derivation and machine learning. Background Art

[0002] Total Organic Carbon (TOC) is a key parameter for evaluating unconventional resources and CO2 geological storage potential. Accurate prediction and quantification of TOC not only helps to evaluate underground resources and determine the geological sweet spots of shale oil, but also helps to indicate the difference in CO2 / shale wettability, thereby advancing CO2 geological storage research. Accurate prediction of TOC is of great significance for the exploration and development of unconventional oil and gas resources and CO2 storage.

[0003] There are generally direct and indirect methods to obtain the TOC content of shale. The direct method is to accurately obtain TOC in shale through rock pyrolysis experiments. This experimental method is often not only time-consuming and labor-intensive, but also costly. Indirect methods such as the ΔLogR method have been cited and improved by many scholars and have achieved good results; however, this method requires artificial setting of the logging curve baseline and the background value of organic carbon content, which is prone to errors and deviations in results, and is greatly affected by geological differences; in the past, scholars analyzed the correlation between shale TOC and multiple logging curves, and selected the logging curve with the best correlation as the prediction parameter for prediction. Some scholars also established a functional relationship with shale TOC based on resistivity logging data and acoustic time difference logging intersection for prediction. The disadvantages of the above methods are that they use a small number of logging curves for analysis, and the parameters in the prediction formula are manually determined standardized data, which has large errors; the establishment of a multivariate linear regression equation is not enough to characterize the complex nonlinear relationship between TOC and various parameters, and between various parameters and various parameters.

[0004] With the development of technology, rapid prediction of TOC can be achieved by analyzing a large amount of logging data based on machine learning algorithms, such as back propagation neural networks and support vector machine algorithms, which can improve accuracy and efficiency compared to traditional methods. Studies have shown that the quantity and quality of data and features determine the upper limit of machine learning, and the model and algorithm should infinitely approach this upper limit. The more input variables, the higher the model accuracy. However, the above technologies still have shortcomings. It is difficult to accurately extract data features, cannot effectively utilize the information contained in the data, and is difficult to process high-latitude data; it is difficult to adapt to the changing geological environment, and lacks flexibility and applicability.

[0005] Therefore, it is necessary to explore a new method that is time-saving, labor-saving, convenient and accurate in predicting the TOC content in shale. Summary of the invention

[0006] The object of the present invention is to provide a shale TOC prediction method and computer program product based on logging data feature derivation and machine learning, which can improve the flexibility, applicability and accuracy of shale TOC prediction.

[0007] The present invention provides a shale TOC prediction method based on logging data feature derivation and machine learning, comprising the following steps: S1: obtaining shale TOC data by using rock pyrolysis experiments; obtaining original logging data by using geological research tools according to the shale TOC data; S2: preprocessing the original logging data to obtain preprocessed logging data; S3: performing feature derivation on the preprocessed logging data to obtain input variables; S4: training a random forest machine learning model by using the shale TOC data and the input variables to obtain a shale TOC prediction model; S5: using the shale TOC prediction model to predict the shale data to be tested to obtain a shale TOC prediction result.

[0008] Further, step S3 specifically includes: S31: according to the preprocessed logging data, using the multivariate linear regression method, to obtain the first characteristic variable; S32: according to the preprocessed logging data, using the Schmoker natural gamma logging method, to obtain the second characteristic variable; S33: according to the preprocessed logging data, using the Schmoker density logging method, to obtain the third characteristic variable; S34: according to the preprocessed logging data, using the natural gamma ray spectrum method, to obtain the fourth characteristic variable; S35: according to the preprocessed logging data, using the ΔLogR method, to obtain the fifth characteristic variable; S36: according to the first characteristic variable, the second characteristic variable, the third characteristic variable, the fourth characteristic variable, the fifth characteristic variable and the preprocessed logging data, to obtain the input variable.

[0009] Further, step S31 specifically includes: according to the pre-processed logging data, using a multivariate linear regression method, obtaining a first characteristic variable, such as formula: , in, is the first characteristic variable, is the natural gamma logging curve, is the logging curve of acoustic wave time difference, is the array induction resistivity logging curve, is the density logging curve, , , , are the linear correlation coefficients between each logging curve and TOC, is a constant.

[0010] Further, step S32 specifically includes: according to the pre-processed logging data, using Schmoker natural gamma logging method, obtaining a second characteristic variable, such as formula: , in, is the second characteristic variable, is the natural gamma logging value, is the linear correlation coefficient between the natural gamma curve and TOC, is a constant.

[0011] Further, step S33 specifically includes: according to the pre-processed logging data, using the Schmoker density logging method, obtaining a third characteristic variable, such as formula: , in, is the third characteristic variable, is the linear correlation coefficient between density curve and TOC, is the density logging response value of rock, is a constant.

[0012] Further, step S34 specifically includes: according to the pre-processed logging data, using natural gamma ray spectroscopy, obtaining a fourth characteristic variable, such as formula: , in, is the fourth characteristic variable, is the radioactivity of uranium in shale, is the radioactivity of potassium in shale, is the linear correlation coefficient between U / K and TOC, is a constant.

[0013] Further, step S35 specifically includes: according to the pre-processed logging data, using the ΔLogR method, obtaining the fifth characteristic variable, such as the formula: , , , in, is the fifth characteristic variable, for and The linear correlation coefficient between is a constant, for and The amplitude difference between the curves, is the resistivity, is the minimum value of resistivity, To calculate the coefficient, is the sound wave time difference, is the minimum value of the sound wave time difference, is the maximum value of resistivity, is the maximum value of the sound wave time difference.

[0014] Furthermore, the above-mentioned input variables include a first characteristic variable, a second characteristic variable, a third characteristic variable, a fourth characteristic variable, a fifth characteristic variable, natural gamma, acoustic wave time difference, array induction resistivity, density, uranium and potassium.

[0015] Furthermore, step S4 specifically includes: S41: obtaining a TOC data set based on the shale TOC data and the input variables; S42: dividing and normalizing the TOC data set according to a preset ratio to obtain a training set and a test set; S43: using the training set and the test set to train a random forest machine learning model to obtain a shale TOC prediction model.

[0016] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned shale TOC prediction method based on logging data feature derivation and machine learning.

[0017] The implementation of the shale TOC prediction method based on logging data feature derivation and machine learning provided by the present invention has the following beneficial effects: The present invention performs preprocessing such as outlier processing and missing value processing on the logging data of the core sampling well, fits the existing logging curve with the experimental analysis TOC data, derives features of the original logging data by methods such as multivariate linear regression, Schmoker natural gamma logging, Schmoker density logging, natural gamma spectroscopy, ΔLogR method, etc., combines the original logging data with the feature derivative data into a new data set, and finally uses the new data set to construct a random forest prediction model, so as to achieve effective prediction of TOC under the condition of small sample data; the present invention uses a method combining a random forest algorithm with the feature derivative of logging data, does not require too many logging data features, is easy to operate, reduces human errors, can more accurately predict TOC for small sample data, greatly reduces the calculation time and research cost of data processing, is simple to calculate and maintains a high accuracy, can improve the flexibility, applicability and accuracy of shale TOC prediction; can also have better prediction results for data with unbalanced and small number of features, and has important significance for reservoir-related physical property analysis, evaluation of reservoir reservoir characteristics and guidance of next step development. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 It is a flow chart of a shale TOC prediction method based on well logging data feature derivation and machine learning provided by the present invention; Figure 2 It is a result diagram of predicting TOC based on original well logging data by using random forest algorithm provided by the present invention; Figure 3 The invention provides a TOC result diagram based on the combination of original well logging data and feature-derived data and predicted by a random forest algorithm; Figure 4 It is a well column sub-diagram comparing the two TOC prediction methods provided by the present invention. DETAILED DESCRIPTION

[0019] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0020] Figure 1 A schematic diagram of a shale TOC prediction method based on well logging data feature derivation and machine learning in this embodiment is shown. In this embodiment, the shale TOC prediction method based on well logging data feature derivation and machine learning includes the following steps: S1: Obtain shale TOC data by using rock pyrolysis experiments; obtain original well logging data by using geological research tools based on the shale TOC data; In an exemplary embodiment, the geological research tool is ResForm software; Specifically, the above raw logging data include natural gamma, acoustic transit time, array induction resistivity, density, uranium and potassium; Specifically, in step S1, the shale TOC (total organic carbon, in wt%) data is obtained through rock pyrolysis experiments, and the parameter data corresponding to TOC are extracted in the geological research tool ResForm software, including natural gamma (abbreviated as GR, in API), acoustic time difference (abbreviated as AC, in μm / s), array induction resistivity (abbreviated as AT120, in Ω·cm), density (abbreviated as DEN, in cm 3 / g), uranium (abbreviated as U, unit is ppm) and potassium (abbreviated as K, unit is %), and the original logging data are obtained; S2: preprocessing the original logging data to obtain preprocessed logging data; Specifically, in step S2, the initially obtained raw logging data is cleaned and processed to remove abnormal values ​​and missing values ​​to ensure the continuity of the data; It should be noted that the above pre-processed logging data include natural gamma, acoustic transit time, array induction resistivity, density, uranium and potassium; S3: performing feature derivation on the preprocessed logging data to obtain input variables; Specifically, step S3 includes: S31: obtaining a first characteristic variable using a multivariate linear regression method according to the preprocessed logging data; Specifically, step S31 includes: according to the preprocessed logging data, using a multivariate linear regression method, obtaining a first characteristic variable, such as the formula: , in, is the first characteristic variable, is the natural gamma logging curve, is the logging curve of acoustic wave time difference, is the array induction resistivity logging curve, is the density logging curve, , , , are the linear correlation coefficients between each logging curve and TOC, is a constant; It should be noted that in step S31, the TOC content can be based on the natural gamma, density, array induction resistivity, acoustic wave time difference and other logging curves to establish a multivariate regression model through SPSS software, that is, ;in , , , are the linear correlation coefficients between each logging curve and TOC, is a constant; this feature is recorded as ; S32: Obtaining a second characteristic variable using the Schmoker natural gamma logging method according to the preprocessed logging data; Specifically, step S32 includes: according to the pre-processed logging data, using the Schmoker natural gamma logging method, obtaining a second characteristic variable, such as the formula: , in, is the second characteristic variable, is the natural gamma logging value, is the linear correlation coefficient between the natural gamma curve and TOC, is a constant; It should be noted that in step S32, since the TOC content in organic-rich shale is higher than that in mudstone, the natural gamma logging value of organic-rich shale is higher than that of mudstone. Therefore, the natural gamma value can be used to identify organic-rich shale, and a correlation formula is established with TOC through the Schmoker natural gamma logging method, that is, ,in is the linear correlation coefficient between the natural gamma curve and TOC, is a constant, denoted as the characteristic ; S33: obtaining a third characteristic variable using the Schmoker density logging method according to the preprocessed logging data; Specifically, step S33 includes: according to the pre-processed logging data, using the Schmoker density logging method, obtaining a third characteristic variable, such as the formula: , in, is the third characteristic variable, is the linear correlation coefficient between density curve and TOC, is the density logging response value of rock, is a constant; It should be noted that in step S33, since the density of organic matter is relatively small compared to the surrounding rock density, the change in formation density has a good correspondence with the change in organic matter abundance. When organic matter replaces the rock skeleton, the density logging response value of the rock will decrease. Therefore, the Schmoker density logging method can be used to establish a correlation with TOC, that is, ;in is the linear correlation coefficient between density curve and TOC, is a constant, denoted as the characteristic ; S34: obtaining a fourth characteristic variable using natural gamma ray spectroscopy according to the preprocessed logging data; Specifically, step S34 includes: according to the pre-processed logging data, using natural gamma ray spectroscopy, obtaining a fourth characteristic variable, such as formula: , in, is the fourth characteristic variable, is the radioactivity of uranium in shale, is the radioactivity of potassium in shale, is the linear correlation coefficient between U / K and TOC, is a constant; It should be noted that in step S34, natural gamma ray spectrum logging mainly reflects the variation characteristics of uranium (U), thorium (Th) and potassium (K) in the formation. The radioactivity in organic-rich shale is significantly higher than that in the surrounding mudstone. The relationship between the natural gamma ray spectrum logging response and the organic matter content shows that the higher radioactive formations are related to the presence of organic matter. Establish a correlation with TOC content, that is ;in is the linear correlation coefficient between U / K and TOC, is a constant, denoted as the characteristic ; S35: obtaining a fifth characteristic variable using a ΔLogR method according to the preprocessed logging data; Specifically, step S35 includes: according to the pre-processed logging data, using the ΔLogR method, obtaining the fifth characteristic variable, such as the formula: , , , in, is the fifth characteristic variable, for and The linear correlation coefficient between is a constant, for and The amplitude difference between the curves, is the resistivity, is the minimum value of resistivity, To calculate the coefficient, is the sound wave time difference, is the minimum value of the sound wave time difference, is the maximum value of resistivity, is the maximum value of the sound wave time difference; It should be noted that in step S35, the ΔLogR method estimates the TOC content by calculating the ΔLogR value based on the cross plot of resistivity (RT) and acoustic time difference (AC). An improved ΔLogR method is cited in the embodiment, which does not require setting a baseline value and an organic carbon content background value, thereby reducing manual errors. The TOC expression obtained by this method is recorded as TOC5; S36: obtaining an input variable according to the first characteristic variable, the second characteristic variable, the third characteristic variable, the fourth characteristic variable, the fifth characteristic variable and the preprocessed logging data; Specifically, the input variables include a first characteristic variable, a second characteristic variable, a third characteristic variable, a fourth characteristic variable, a fifth characteristic variable, natural gamma, acoustic wave time difference, array induction resistivity, density, uranium and potassium; Specifically, in step S36, finally, after deriving the logging data characteristics, a total of 11 input variables are obtained: , , , , , , , , , , ; S4: Using the shale TOC data and the input variables to train a random forest machine learning model to obtain a shale TOC prediction model; Specifically, step S4 includes: S41: Obtaining a TOC data set according to the shale TOC data and the input variables; In an exemplary embodiment, for step S41, a random forest machine learning model is established through Python software. First, the environment variables are cleared and data is imported; the sorted data is imported into the random forest program, and the parameters GR, AC, AT120, DEN, U, K, TOC1, TOC2, TOC3, TOC4, and TOC5 are set as input variables, and TOC is set as an output variable, and 11 input variables at the same depth are specified. And the corresponding TOC content A set of data is formed, where n is 1 to 11, representing different parameters, and i represents different depths. All data are integrated in groups according to different depths to obtain the TOC data set; S42: dividing and normalizing the TOC data set according to a preset ratio to obtain a training set and a test set; Specifically, in step S42, the preliminarily processed data is imported into the program, the data is randomly rearranged in groups, the training set and the test set are divided according to a preset ratio of 80%:20%, and the input and output data of the training set and the test set are normalized; S43: Using the training set and the test set to train the random forest machine learning model to obtain a shale TOC prediction model; Specifically, in step S43, the parameters in the random forest machine learning model are adjusted during training: the number of decision trees, random seeds, minimum number of leaves, and maximum depth; after debugging, the model is used to train and test the sorted data set, and the training is repeated to fit the TOC predicted values ​​in the training set and the TOC predicted values ​​in the test set with the corresponding TOC measured values ​​until the fitting results of the training set and the test set achieve a suitable effect, thereby obtaining a shale TOC prediction model; S5: using the shale TOC prediction model to predict the shale data to be tested, and obtaining a shale TOC prediction result; Specifically, in step S5, the trained model is used to predict new data, the data to be predicted is imported into the model for verification, and the TOC prediction result is output.

[0021] In an exemplary embodiment, a shale TOC prediction method based on well logging data feature derivation and machine learning includes the following steps: Step 1: Obtain shale TOC data through rock pyrolysis experiments, extract parameter data corresponding to TOC in ResForm software, including natural gamma (GR), acoustic time difference (AC), array induction resistivity (AT120), density (DEN), uranium (U) and potassium (K), and obtain original logging data; Step 2: Clean and process the raw data obtained initially, remove outliers and missing values, and ensure the continuity of the data; Step 3: The original logging data were characterized by multiple linear regression, Schmoker natural gamma logging, Schmoker density logging, natural gamma spectroscopy, ΔLogR and other methods, and the relationship between the original logging data and the experimental TOC was established, which were recorded as TOC1, TOC2, TOC3, TOC4 and TOC5 respectively. After the logging data characteristics were derived, a total of 11 input variables were obtained: GR, AC, AT120, DEN, U, K, TOC1, TOC2, TOC3, TOC4 and TOC5. Step 4: Use Python software to build a random forest machine learning model. First, clear the environment variables and import the data. Import the sorted data into the random forest program, set the parameters GR, AC, AT120, DEN, U, K, TOC1, TOC2, TOC3, TOC4, and TOC5 as input variables, set TOC as the output variable, and specify 11 input variables at the same depth. And the corresponding TOC content Construct a set of data, where n is 1 to 11, representing different parameters, and i represents different depths. All data are integrated in groups according to different depths; import the preliminarily processed data into the program, randomly rearrange the data in groups, divide them into training sets (80%) and test sets (20%), and normalize the input and output data of the training sets and test sets; adjust the parameters in the training model: the number of decision trees, random seeds, minimum number of leaves, and maximum depth; after debugging, use the model to train and test the sorted data set, and train repeatedly to fit the TOC predicted values ​​in the training set and the TOC predicted values ​​in the test set with the corresponding TOC measured values ​​until the fitting results of the training set and the test set reach a suitable effect; Step 5: After training the model, predict new data, import the data to be predicted into the model for verification, and output the predicted TOC to get the result; The predicted values ​​were fitted with the actual values, and the TOC prediction accuracy was improved from 0.8292 to 0.9207. Figure 2 It is the TOC result diagram predicted by random forest algorithm based on the original logging data; Figure 3 It is based on the combination of original logging data and feature-derived data, and predicts the TOC result map through the random forest algorithm; Figure 4 It is a well column sub-diagram comparing two TOC prediction methods; Method 1 is predicted by original logging data, and Method 2 is predicted by combining original logging data with feature-derived data. It can be seen that the TOC model has higher prediction accuracy after data feature-derived processing.

[0022] This embodiment provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned shale TOC prediction method based on logging data feature derivation and machine learning.

[0023] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A shale TOC prediction method based on well logging data feature derivation and machine learning, characterized in that: The following steps are involved: S1: Obtain shale TOC data by using rock pyrolysis experiments; obtain original well logging data by using geological research tools based on the shale TOC data; S2: preprocessing the original logging data to obtain preprocessed logging data; S3: performing feature derivation on the preprocessed logging data to obtain input variables; S4: Using the shale TOC data and the input variables to train a random forest machine learning model to obtain a shale TOC prediction model; S5: Use the shale TOC prediction model to predict the shale data to be tested to obtain a shale TOC prediction result.

2. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 1 is characterized in that: Step S3 specifically includes: S31: obtaining a first characteristic variable using a multivariate linear regression method according to the preprocessed logging data; S32: Obtaining a second characteristic variable using the Schmoker natural gamma logging method according to the preprocessed logging data; S33: obtaining a third characteristic variable using the Schmoker density logging method according to the preprocessed logging data; S34: obtaining a fourth characteristic variable using natural gamma ray spectroscopy according to the preprocessed logging data; S35: obtaining a fifth characteristic variable using a ΔLogR method according to the preprocessed logging data; S36: Obtain input variables according to the first characteristic variable, the second characteristic variable, the third characteristic variable, the fourth characteristic variable, the fifth characteristic variable and the preprocessed logging data.

3. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2 is characterized in that: Step S31 specifically includes: according to the pre-processed logging data, using a multivariate linear regression method, obtaining a first characteristic variable, such as the formula: ， in, is the first characteristic variable, is the natural gamma logging curve, is the logging curve of acoustic wave time difference, is the array induction resistivity logging curve, is the density logging curve, , , , are the linear correlation coefficients between each logging curve and TOC, is a constant.

4. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2, characterized in that: Step S32 specifically includes: according to the pre-processed logging data, using the Schmoker natural gamma logging method, obtaining a second characteristic variable, such as the formula: ， in, is the second characteristic variable, is the natural gamma logging value, is the linear correlation coefficient between the natural gamma curve and TOC, is a constant.

5. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2, characterized in that: Step S33 specifically includes: according to the pre-processed logging data, using the Schmoker density logging method, obtaining a third characteristic variable, such as the formula: ， in, is the third characteristic variable, is the linear correlation coefficient between density curve and TOC, is the density logging response value of rock, is a constant.

6. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2, characterized in that: Step S34 specifically includes: according to the pre-processed logging data, using natural gamma ray spectroscopy, obtaining a fourth characteristic variable, such as formula: ， in, is the fourth characteristic variable, is the radioactivity of uranium in shale, is the radioactivity of potassium in shale, is the linear correlation coefficient between U / K and TOC, is a constant.

7. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2, characterized in that: Step S35 specifically includes: according to the pre-processed logging data, using the ΔLogR method, obtaining the fifth characteristic variable, such as the formula: ， ， ， in, is the fifth characteristic variable, for and The amplitude difference between the curves, for and The linear correlation coefficient between is a constant, is the resistivity, is the minimum value of resistivity, To calculate the coefficient, is the sound wave time difference, is the minimum value of the sound wave time difference, is the maximum value of resistivity, is the maximum value of the sound wave time difference.

8. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 2, characterized in that: The input variables include a first characteristic variable, a second characteristic variable, a third characteristic variable, a fourth characteristic variable, a fifth characteristic variable, natural gamma, acoustic wave time difference, array induction resistivity, density, uranium and potassium.

9. The shale TOC prediction method based on well logging data feature derivation and machine learning according to claim 1, characterized in that: Step S4 specifically includes: S41: Obtaining a TOC data set according to the shale TOC data and the input variables; S42: dividing and normalizing the TOC data set according to a preset ratio to obtain a training set and a test set; S43: Using the training set and the test set to train the random forest machine learning model, to obtain a shale TOC prediction model.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the shale TOC prediction method based on logging data feature derivation and machine learning as described in any one of claims 1 to 9 are implemented.

Images