A shale lithology identification method based on prior Bayesian cascade forest

Abstract

The invention discloses a shale lithology identification method based on a priori Bayesian cascade forest, and relates to the field of rock. The shale lithology identification method based on a priori Bayesian cascade forest comprises the following steps: Step 1, firstly establishing a mineral pre-training model data set of terrestrial shale, then taking the logging data in the data set as features and the mineral data as labels, using a Bayesian extreme gradient boosting regressor for training, and saving the mineral pre-training model; compared with the prior art, the invention has the following beneficial effects: the invention takes into account the class imbalance problem of logging data, uses adaptive multi-target group fusion to resample the data set, uses X-ray diffraction whole-rock mineral analysis data, generates features with the help of a Bayesian gradient boosting regressor, and then uses a Bayesian gradient boosting classifier to identify shale lithology, which has high accuracy and low cost, and provides a reliable solution for shale lithology identification.

Description

Technical Field

[0001] The invention relates to the field of rocks, and in particular to a shale lithology identification method based on a priori Bayesian cascade forest. Background Art

[0002] Continental shale reservoirs are characterized by diverse lithology, frequent vertical lithology changes, and strong reservoir heterogeneity. Accurate identification of shale lithology is of great significance for shale reservoir evaluation and development. Laboratory core analysis is the most accurate lithology identification method. Scanning electron microscopy and X-ray diffraction whole-rock mineral analysis (XRD) are both laboratory lithology identification methods. Although laboratory lithology identification methods such as lithology identification are highly accurate, they are expensive and difficult to obtain, making them difficult to apply to the entire formation. Therefore, well logging lithology identification has begun to emerge. In the field of well logging lithology identification, the intersection plot method is a classic method for identifying lithology, but the identification effect is not ideal; after the introduction of machine learning algorithms, people have proposed different methods such as artificial neural network-based identification models and BP neural network models to identify shale lithology or lithology, but due to the class imbalance problem when classifying well data, the accuracy is low and the final identification effect needs to be improved.

[0003] In summary, among the existing shale lithology identification methods, there are few methods with low cost and high accuracy, which need to be improved. Summary of the invention

[0004] The purpose of the present invention is to provide a shale lithology identification method based on a priori Bayesian cascade forest to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A shale lithology identification method based on a priori Bayesian cascade forest includes the following steps:

[0007] Step 1: First, establish a mineral pre-training model dataset for continental shale, then use the well logging data in the dataset as features and the mineral data as labels, use the Bayesian extreme gradient boosting regressor for training, and save the mineral pre-training model;

[0008] Step 2, establishing a lithology prediction model training data set for continental shale, and resampling the lithology prediction model training data set using an adaptive multi-objective group fusion method;

[0009] Step 3: Based on the resampled lithology prediction model training data set, the Bayesian extreme gradient boosting classifier is used for training to generate a lithology prediction model. The data to be identified is then input into the mineral pre-training model to generate mineral characteristics, which are then input into the prediction model to obtain the final lithology prediction result.

[0010] As a further solution of the present invention: Step 1 comprises:

[0011] Step 11, establishing a continental shale mineral pre-training model data set including well logging data and mineral data; wherein the well logging data is the well logging data of different continental shale rock samples, and the mineral data is XRD mineral data;

[0012] Step 12, using the logging data of different continental shale rock samples as features and the XRD mineral data as labels, the Bayesian extreme gradient boosting regressor is used for training to obtain the mineral pre-training model and save it.

[0013] As a further solution of the present invention: in step 11, the XRD mineral data includes the mass content of minerals, and the minerals include quartz, potassium feldspar, plagioclase, calcite, ankerite, dolomite, siderite, pyrite, and clay. The logging data of different continental shale rock samples include acoustic time difference (AC), neutron logging (CNL), density logging (DEN), natural gamma logging (GR), volume photoelectric cross-section absorption index (PE), and deep lateral resistivity (RLLD).

[0014] As a further solution of the present invention: in step 12, the complexity of continental shale is dealt with by using the extreme gradient boosting method after the Bayesian optimization parameter search process through the Bayesian extreme gradient boosting regressor;

[0015] Objective Function of Extreme Gradient Boosting (XGBoost) for:

[0016] (1);

[0017] in, is the loss function, is the total node value, is the current node value, is the number of samples, is the number of iterations, For the The first derivative of the predicted value of the iteration, For the The second derivative of the predicted value of the iteration, Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree;

[0018] Since XGBoost has target evaluation indicators Therefore, we need to use Bayesian optimization extreme gradient boosting (XGBoost) to find the optimal learner parameters.

[0019] (2);

[0020] in represents a parameter vector, Pa represents a set of parameter vectors;

[0021] When determining the optimization direction, the expected improvement method is used as the acquisition function, that is:

[0022] (3);

[0023] in represents the model index, represents the model indicator threshold, Indicated in The degree of improvement of model indicators under parameter vector, Represents the probability distribution model based on the existing parameter adjustment history;

[0024] At this point, the next parameter vector to be searched is for

[0025]

[0026] In this process, the indicator threshold This is achieved through Gaussian process regression; at this point, the mineral pre-training data set is put into the Bayesian extreme gradient boosting regressor for training to obtain the mineral pre-training model.

[0027] As a further solution of the present invention: Step 2 comprises:

[0028] Step 21, establishing a lithology prediction model training data set for continental shale including lithology information and well logging data;

[0029] Step 22, resample the training set using the adaptive multi-target group fusion method. The adaptive multi-target group fusion method (AMSCO) constructs two subgroups, one of which optimizes the majority class through group instance selection (SIS), and the other optimizes the minority class through the oversampling technique of synthetic minority class instances (OSMOTE); the oversampling technique (OSMOTE) inserts the generated samples into any or K minority class nearest neighbor segments in the data space (K is also called the Kappa value), and optimizes through particle swarm optimization (PSO). During the particle optimization process, three update criteria need to be met: the accuracy and Kappa value of the new particle must be better than the existing particle; the current Kappa threshold of the new particle must be better than the Kappa value of the old particle.

[0030] As a further solution of the present invention: in step 21, the training data set of the lithology prediction model of continental shale includes lithology information and logging data of different continental shale rock samples, and the lithology information of different continental shale rock samples is derived from the thin section identification results of the rock samples; the logging data of different continental shale rock samples include acoustic time difference (AC), neutron logging (CNL), density logging (DEN), natural gamma logging (GR), volume photoelectric cross-section absorption index (PE), and deep lateral resistivity (RLLD).

[0031] As a further solution of the present invention: in step 22, the accuracy and Kappa value of the new particle that are better than the existing particle can be replaced by one of the accuracy and Kappa value of the new particle that is better than the existing particle.

[0032] As a further solution of the present invention: step 3 comprises:

[0033] Step 31, input the resampled lithology prediction model training data set into the mineral pre-training model obtained in step 1 to generate new continental shale mineral results; then, the well logging data and the newly generated mineral results in the resampled lithology prediction model training data set are used as features, and the lithology is used as a label, and the Bayesian extreme gradient boosting classifier is used again for training to generate a lithology prediction model;

[0034] Step 32, input the data to be identified into the continental shale mineral pre-training model obtained in step 1 to generate new mineral results, and then input the new mineral results and the logging data to be identified into the lithology prediction model to obtain the final lithology prediction results.

[0035] Compared with the prior art, the present invention has the following beneficial effects: the present invention takes into account the class imbalance problem of logging data, uses adaptive multi-objective group fusion (AMSCO) to resample the data set, uses X-ray diffraction whole-rock mineral analysis (XRD) data, generates features with the help of a Bayesian gradient boosting regressor, and then uses a Bayesian gradient boosting classifier to identify shale lithology, which has high accuracy and low cost, and provides a reliable solution for shale lithology identification;

[0036] The mineral pre-training model in the present invention maps the discrete laboratory mineral content measurement data with the continuous well data through the Bayesian extreme gradient boosting method. This association adds a new dimension to the well data, forming continuous well mineral content data and increasing the information of the data set.

[0037] The Bayesian extreme gradient boosting used in the present invention is a tree ensemble algorithm based on the gradient boosting framework, which uses decision trees as weak learners. Compared with other machine learning methods such as neural networks, it has a simpler structure and a more interpretable model.

[0038] The Bayesian extreme gradient boosting method used in the present invention uses a gradient boosting framework based on residual learning ideas with the Bayesian optimization theory of prior constraints and posteriori, has the characteristics of adaptively adjusting hyperparameters, is easier to apply, and can achieve the optimal effect within a fixed time cost;

[0039] The present invention uses adaptive multi-target group fusion to adaptively improve the imbalance problem of data in continental shale wells and improve the accuracy of shale lithology identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flow chart of a shale lithology identification method based on prior Bayesian cascade forest.

[0041] Figure 2 This is the intersection matrix diagram of continental shale logging data.

[0042] Figure 3 This is the effect diagram of resampling the foreland shale logging data.

[0043] Figure 4 This is the effect diagram of continental shale logging data after resampling.

[0044] Figure 5 Prior Bayesian cascade forests applied to continental shale stratigraphic detection map. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0046] See also Figure 1 , a shale lithology identification method based on a priori Bayesian cascade forest, comprising the following steps:

[0047] Step 1: First, establish a mineral pre-training model dataset for continental shale, then use the well logging data in the dataset as features and the mineral data as labels, use the Bayesian extreme gradient boosting regressor for training, and save the mineral pre-training model;

[0048] Step 2, establishing a lithology prediction model training data set for continental shale, and resampling the lithology prediction model training data set using an adaptive multi-objective group fusion method;

[0049] Step 3: Based on the resampled lithology prediction model training data set, the Bayesian extreme gradient boosting classifier is used for training to generate a lithology prediction model. The data to be identified is then input into the mineral pre-training model to generate mineral characteristics, which are then input into the prediction model to obtain the final lithology prediction result.

[0050] In this example: See Figure 1 , step 1 includes:

[0051] Step 11, establish a continental shale mineral pre-training model data set including well logging data and mineral data, use core calibration logging to find the well logging data corresponding to the discrete core mineral data sampling depth, extract natural gamma, density, sonic, neutron and other well logging data points point by point and construct a data set; the well logging data is the well logging data of different continental shale rock samples, and the mineral data is XRD mineral data;

[0052] Step 12, using the logging data of different continental shale rock samples as features and the XRD mineral data as labels, the Bayesian extreme gradient boosting regressor is used for training to obtain the mineral pre-training model and save it.

[0053] In this example: See Figure 1In step 11, the XRD mineral data includes the mass content of minerals, and the minerals include quartz, potassium feldspar, plagioclase, calcite, ankerite, dolomite, siderite, pyrite, and clay. The logging data of different continental shale rock samples include acoustic transit time (AC), neutron logging (CNL), density logging (DEN), natural gamma logging (GR), volume photoelectric cross-section absorption index (PE), and deep lateral resistivity (RLLD).

[0054] In this example: See Figure 1 and Figure 2 In step 12, the well logging data is a typical low-dimensional discrete data. At the same time, because shale is usually mixed with other lithologies and the distribution of different shales is also complex, the well logging data set under complex lithology has a typical class imbalance characteristic, such as Figure 2 As shown, it is necessary to combine the gradient boosting method which has advantages in the well logging lithology identification scenario. However, the common gradient boosting method is difficult to cope with the complexity of continental shale. Therefore, the Bayesian extreme gradient boosting regressor is used to use the extreme gradient boosting method after the Bayesian optimization parameter search process to cope with the complexity of continental shale.

[0055] Objective Function of Extreme Gradient Boosting (XGBoost) for:

[0056] (1);

[0057] in, is the loss function, is the total node value, is the current node value, is the number of samples, is the number of iterations, For the The first derivative of the predicted value of the iteration, For the The second derivative of the predicted value of the iteration, Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree;

[0058] Since XGBoost has target evaluation indicators Therefore, we need to use Bayesian optimization extreme gradient boosting (XGBoost) to find the optimal learner parameters.

[0059] (2);

[0060] in represents a parameter vector, Pa represents a set of parameter vectors;

[0061] When determining the optimization direction, the expected improvement method is used as the acquisition function, that is:

[0062] (3);

[0063] in represents the model index, represents the model indicator threshold, Indicated in The degree of improvement of model indicators under parameter vector, Represents the probability distribution model based on the existing parameter adjustment history;

[0064] At this point, the next parameter vector to be searched is for

[0065]

[0066] In this process, the indicator threshold This is achieved through Gaussian process regression;

[0067] Assume that the target evaluation index and target parameter obey Gaussian distribution and set their mean to 0. This distribution is used as the prior distribution. Further, the posterior distribution of the target evaluation index and target parameter is obtained based on the observed points. As the number of observation points increases, the target evaluation index is continuously updated, and the best index of the current observation result is used as the index threshold. . Then continue searching for better parameters until the iteration stops.

[0068] At this point, the mineral pre-training model can be obtained by training the mineral pre-training dataset into the Bayesian extreme gradient boosting regressor.

[0069] In this example: See Figure 1 , step 2 includes:

[0070] Step 21, establish a lithology prediction model training data set for continental shale containing lithology information and logging data, use the core thin section identification calibration logging data method to find the logging data corresponding to the sampling depth of discrete core lithology data, extract natural gamma, density, sonic, neutron and other logging data points point by point and construct a data set, the lithology label comes from the core thin section identification, the discrete lithology label is matched with the logging data through the core calibration logging method, and the data set is constructed;

[0071] Step 22, resample the training set using an adaptive multi-target group fusion method,

[0072] Since the information of the majority class in an unbalanced data set far exceeds that of the minority class, the classifier is prone to overfitting the majority class. Conventional well logging data sets are usually unbalanced data sets. Therefore, an adaptive multi-objective group fusion method (AMSCO) is used to deal with the unbalanced classification problem.

[0073] The AMSCO method proposes an adaptive rebalancing model, called population fusion, which focuses on finding the best combination between majority and minority class data during the rebalancing process. The core idea is to decompose the original dataset and then recombine it using only qualified samples selected by the two population optimization processes. During the rebalancing process, the parameters are adjusted automatically and adaptively.

[0074] The core optimizer of AMSCO is particle swarm optimization (PSO). The PSO algorithm simulates the behavior of birds foraging and is a classic swarm intelligence algorithm. Compared with other intelligent algorithms, it is easy to implement, has a faster convergence speed and fewer parameters. Since the objective function and constraints of the PSO algorithm are relatively simple, it is widely used in various fields.

[0075] During the optimization process of PSO, the initial dataset is used as the current dataset. The candidate classifier is used to verify the quality of the current dataset. The current dataset will undergo two parallel group optimizations to change the sample distribution in the best way until the performance of the candidate classifier reaches the threshold. Although the two candidate solutions are different in nature, a dataset instance close to the size of the initial dataset is generated by selectively merging information from the best dataset instance into one instance. In the iterative loop, this dataset instance will be used as the current dataset and then checked by the evaluation index. If the evaluation index is less than the threshold, the dataset will be divided into two parallel group optimization processes again.

[0076] AMSCO couples these two parallel optimizations together as a unified iterative process. It gradually enhances the data mix from the two group optimizations through iterations until a high-quality dataset is generated. The various subgroups divide and search the space in parallel and share the best information, enabling the best solution to be found in the shortest time.

[0077] The adaptive multi-target group fusion method (AMSCO) constructs two subgroups, one of which optimizes the majority class through group instance selection (SIS), which determines the best group instance generated by comparing the results of multiple groups of group instances obtained after deletion, and the other optimizes the minority class through the oversampling technology of synthetic minority class instances (OSMOTE), that is, the minority class instances are interpolated through the distance strategy to expand the sample size of the minority class instances; the oversampling technology (OSMOTE) inserts the generated samples into any or K minority class nearest neighbor segments in the data space (K is also called the Kappa value), and optimizes them through particle swarm optimization (PSO). During the particle optimization process, three update criteria need to be met: the accuracy and Kappa value of the new particles must be better than the existing particles; the current Kappa threshold of the new particles must be better than the Kappa value of the old particles.

[0078] See also Figure 3 and Figure 4 , based on the above AMSCO method, the lithology prediction model training data set of continental shale was resampled, and clay felsic shale samples (red scattered points) and dolomite samples (blue scattered points) were used as representatives to prove the effectiveness of the AMSCO method ( Figure 3 To test accuracy before resampling, Figure 4 is the test accuracy after resampling).

[0079] In this example: See Figure 1 In step 21, the lithology prediction model training data set of continental shale includes lithology information and logging data of different continental shale rock samples. The lithology information of different continental shale rock samples comes from the thin section identification results of the rock samples; the logging data of different continental shale rock samples include acoustic transit time (AC), neutron logging (CNL), density logging (DEN), natural gamma logging (GR), volume photoelectric cross-section absorption index (PE), and deep lateral resistivity (RLLD).

[0080] The lithology prediction model training data set is consistent with the logging curve of the above-mentioned mineral pre-training model data set.

[0081] In this example: See Figure 1 In step 22, the accuracy and Kappa value of the new particle are both better than the existing particle and can be replaced by one of the accuracy and Kappa value of the new particle that is better than the existing particle.

[0082] In this example: See Figure 1 , step 3 includes:

[0083] Step 31, input the resampled lithology prediction model training data set into the mineral pre-training model obtained in step 1 to generate new continental shale mineral results; then, the well logging data and the newly generated mineral results in the resampled lithology prediction model training data set are used as features, and the lithology is used as a label, and the Bayesian extreme gradient boosting classifier is used again for training to generate a lithology prediction model;

[0084] Step 32, input the data to be identified into the continental shale mineral pre-training model obtained in step 1 to generate new mineral results, and then input the new mineral results and the logging data to be identified into the lithology prediction model to obtain the final lithology prediction results.

[0085] See also Figure 5 In order to verify the applicability of the model, the prior Bayesian cascade forest was applied to a continental shale formation. The sixth discrete data track in the figure is the result of coring identification of the wellbore, and the seventh track is the lithology result of the continuous formation predicted by the present invention. The lithology prediction of most layers is accurate, and the lithology gap of the uncored layers is filled. The results show that the present invention is superior to other methods and provides a reliable solution for shale lithology identification.

[0086] In the present invention, the mineral pre-training model maps the discrete laboratory mineral content measurement data with the continuous well data through the Bayesian extreme gradient boosting method. This association adds a new dimension to the well data, forming continuous well mineral content data and increasing the information of the data set.

[0087] The Bayesian extreme gradient boosting used in the present invention is a tree ensemble algorithm based on the gradient boosting framework, which uses decision trees as weak learners. Compared with other machine learning methods such as neural networks, it has a simpler structure and a more interpretable model.

[0088] The Bayesian extreme gradient boosting method used in the present invention uses a gradient boosting framework based on residual learning ideas with the Bayesian optimization theory of prior constraints and posteriori, has the characteristics of adaptively adjusting hyperparameters, is easier to apply, and can achieve the optimal effect within a fixed time cost;

[0089] The present invention uses adaptive multi-target group fusion to adaptively improve the imbalance problem of data in continental shale wells and improve the accuracy of shale lithology identification.

[0090] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered as exemplary and non-restrictive in all respects.

[0091] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A shale lithology identification method based on a priori Bayesian cascade forest, characterized in that: The shale lithology identification method based on prior Bayesian cascade forest includes the following steps: Step 1: First, establish a mineral pre-training model dataset for continental shale, then use the well logging data in the dataset as features and the mineral data as labels, use the Bayesian extreme gradient boosting regressor for training, and save the mineral pre-training model; Step 2, establishing a lithology prediction model training data set for continental shale, and resampling the lithology prediction model training data set using an adaptive multi-objective group fusion method; Step 3: Based on the resampled lithology prediction model training data set, the Bayesian extreme gradient boosting classifier is used for training to generate a lithology prediction model. The data to be identified is then input into the mineral pre-training model to generate mineral characteristics, which are then input into the prediction model to obtain the final lithology prediction result.

2. The shale lithology identification method based on prior Bayesian cascade forest according to claim 1 is characterized in that: Step 1 includes: Step 11, establishing a continental shale mineral pre-training model data set including well logging data and mineral data; wherein the well logging data is the well logging data of different continental shale rock samples, and the mineral data is XRD mineral data; Step 12, using the logging data of different continental shale rock samples as features and the XRD mineral data as labels, the Bayesian extreme gradient boosting regressor is used for training to obtain the mineral pre-training model and save it.

3. The shale lithology identification method based on prior Bayesian cascade forest according to claim 2 is characterized in that: In step 11, the XRD mineral data includes the mass content of minerals, and the minerals include quartz, potassium feldspar, plagioclase, calcite, ankerite, dolomite, siderite, pyrite, and clay. The logging data of different continental shale rock samples include sonic time difference, neutron logging, density logging, natural gamma logging, volume photoelectric cross-section absorption index, and deep lateral resistivity.

4. The shale lithology identification method based on prior Bayesian cascade forest according to claim 2 or 3, characterized in that: In step 12, the complexity of continental shale is dealt with by using the extreme gradient boosting method after the Bayesian optimization parameter search process through the Bayesian extreme gradient boosting regressor; Objective function of extreme gradient boosting for: （1）； in, is the loss function, is the total node value, is the current node value, is the number of samples, is the number of iterations, For the The first derivative of the predicted value of the iteration, For the The second derivative of the predicted value of the iteration, Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree; Since extreme gradient boosting has target evaluation indicators Therefore, it is necessary to use Bayesian optimization to find the optimal learner parameters through extreme gradient boosting. （2）； in represents a parameter vector, Pa represents a set of parameter vectors; When determining the optimization direction, the expected improvement method is used as the acquisition equation, that is: （3）； in represents the model index, represents the model indicator threshold, Indicated in The degree of improvement of model indicators under parameter vector, Represents the probability distribution model based on the existing parameter adjustment history; At this point, the next parameter vector to be searched is for: ； In this process, the indicator threshold This is achieved through Gaussian process regression; at this point, the mineral pre-training data set is put into the Bayesian extreme gradient boosting regressor for training to obtain the mineral pre-training model.

5. The shale lithology identification method based on prior Bayesian cascade forest according to claim 1 is characterized in that: Step 2 includes: Step 21, establishing a lithology prediction model training data set for continental shale including lithology information and well logging data; Step 22, resample the training set using an adaptive multi-target group fusion method. The adaptive multi-target group fusion method constructs two subgroups, one of which optimizes the majority class through group instance selection, and the other optimizes the minority class through an oversampling technique that synthesizes minority class instances; the oversampling technique inserts the generated samples into any or K minority class nearest neighbor segments in the data space and optimizes them through particle swarm optimization. During the particle optimization process, three update criteria need to be met: the accuracy and Kappa value of the new particles must be better than the existing particles; the current Kappa threshold of the new particles must be better than the Kappa value of the old particles.

6. The shale lithology identification method based on prior Bayesian cascade forest according to claim 5 is characterized in that: In step 21, the training data set of the lithology prediction model of continental shale includes lithology information and logging data of different continental shale rock samples. The lithology information of different continental shale rock samples comes from the thin section identification results of the rock samples; the logging data of different continental shale rock samples include sonic time difference, neutron logging, density logging, natural gamma logging, volume photoelectric cross-section absorption index, and deep lateral resistivity.

7. The shale lithology identification method based on prior Bayesian cascade forest according to claim 5 or 6, characterized in that: In step 22, the accuracy and Kappa value of the new particle that are better than the existing particle can be replaced by one of the accuracy and Kappa value of the new particle that are better than the existing particle.

8. The shale lithology identification method based on prior Bayesian cascade forest according to claim 1 is characterized in that: Step 3 includes: Step 31, input the resampled lithology prediction model training data set into the mineral pre-training model obtained in step 1 to generate new continental shale mineral results; then, the well logging data and the newly generated mineral results in the resampled lithology prediction model training data set are used as features, and the lithology is used as a label, and the Bayesian extreme gradient boosting classifier is used again for training to generate a lithology prediction model; Step 32, input the data to be identified into the continental shale mineral pre-training model obtained in step 1 to generate new mineral results, and then input the new mineral results and the logging data to be identified into the lithology prediction model to obtain the final lithology prediction results.

Images