Lithofacies identification method for sea-land transition phase shale

Abstract

The invention discloses a lithofacies identification method for sea-land transition phase shale, and relates to the technical field of lithofacies identification. Comprising the following steps: acquiring historical logging curve data of sea-land transition phase shale and corresponding historical lithofacies type information as a training data set; the method comprises the steps that at least two different types of machine learning models serve as base learners, a logistic regression model Logistic Regression serves as a meta classifier, and a stacking model is constructed; training the stack model through the training data set to obtain a trained stack model; and in the lithofacies identification process of the sea-land transition phase shale, inputting the sea-land transition phase shale to be identified into the trained stacking model to obtain a lithofacies identification result. According to the method, the lithology of the sea-land transition phase shale can be better captured, and the robustness and stability of lithofacies identification and classification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithofacies identification, and in particular to a lithofacies identification method for marine-continental transitional shale. Background Art

[0002] Lithofacies refers to the rock types and combinations formed in a specific sedimentary environment, encompassing rock properties, structure, and texture. Transitional shales are characterized by complex and rapidly changing lithofacies. Identifying the differences and distribution patterns among different shale lithofacies can provide an effective basis for exploring shale gas sweet spots.

[0003] Currently, shale lithofacies classification is mostly based on coring and testing data, and there is a lack of intelligent, quantitative identification methods for different shale lithofacies based on well logs. The development of machine learning technology, which automatically identifies and classifies shale lithofacies by learning patterns from historical data, has provided a new solution for lithofacies classification.

[0004] Although machine learning methods have achieved high accuracy in lithofacies classification, the lithology of transitional shales is complex and varied, with various types of source-reservoir combinations, such as sandy interbedded types, limy mixed deposits, and organic-rich mudstone stratification. Different types of shales differ in their logging response characteristics, and the accuracy of lithofacies identification of transitional shales using existing machine learning methods is low. Summary of the Invention

[0005] Based on this, it is necessary to provide a lithofacies identification method for marine-continental transitional shale to address the above technical issues.

[0006] An embodiment of the present invention provides a lithofacies identification method for marine-continental transitional shale, comprising: Obtain historical well logging data and corresponding historical lithofacies type information of marine-continental transitional shale as a training dataset; Use at least two different types of machine learning models as base learners and the logistic regression model LogisticRegression as a meta-classifier to build a stacked model; The training data set is input into the stacking model, and each base learner is trained using the training data set to obtain multiple trained base learners. Each trained base learner is used to predict the training data set to obtain the lithofacies prediction results of the marine-continental transitional shale. The lithofacies prediction results of each trained base learner are used as new features to form a new feature space. The meta-classifier is trained using the new feature space to obtain a trained meta-classifier, thereby obtaining a trained stacking model. In the lithofacies identification process of the transitional shale, the well logging curve data of the transitional shale to be identified is input into the trained stacking model to obtain the lithofacies type identification result.

[0007] Optionally, the at least two different types of machine learning models include: an XGBoost model, a Random Forest model, a CatBoost model, and a LightGBM model.

[0008] Optionally, use the XGBoost model as a base learner, specifically including: Determine the search range of the hyperparameters of the XGBoost model; the hyperparameters of the XGBoost model include: learning rate, maximum tree depth, number of weak classifiers, sampling ratio, and the ratio of features used in each tree; During the grid search process, the value of each hyperparameter of the XGBoost model is determined to form different hyperparameter combinations; Perform cross-validation on each hyperparameter combination to maximize the accuracy of the XGBoost model and obtain the optimal hyperparameter combination; The XGBoost model corresponding to the optimal hyperparameter combination is used as the XGBoost model after hyperparameter optimization, and the XGBoost model after hyperparameter optimization is used as the base learner.

[0009] Optionally, a Random Forest model is used as a base learner, specifically including: Determine the search range for the hyperparameters of the Random Forest model; the hyperparameters of the Random Forest model include: the number of decision trees, the maximum tree depth, the minimum number of split samples, the minimum number of samples per leaf node, and whether to perform bootstrap sampling; During the grid search process, the effects of different combinations of hyperparameters on the performance of the Random Forest model are calculated to determine the hyperparameter combination with the best accuracy. The Random Forest model corresponding to the hyperparameter combination with the best accuracy is used as the Random Forest model after hyperparameter optimization, and the Random Forest model after hyperparameter optimization is used as the base learner.

[0010] Optionally, a CatBoost model is used as a base learner, specifically including: Determine the search range of the hyperparameters of the CatBoost model; the hyperparameters of the CatBoost model include: learning rate, tree depth and number of iterations; During the grid search process, each hyperparameter combination of the CatBoost model in the hyperparameter grid is traversed; Evaluate the performance of the CatBoost model under each hyperparameter combination through cross-validation to determine the hyperparameter combination with the best performance; The CatBoost model corresponding to the best hyperparameter combination is used as the CatBoost model after hyperparameter optimization, and the CatBoost model after hyperparameter optimization is used as the base learner.

[0011] Optionally, historical well logging data and corresponding historical lithofacies type information of marine-continental transitional shale are obtained as a training data set, specifically including: Obtain input data files including historical well logging data of gamma ray GR, acoustic travel time AC, neutron porosity CNL, density DEN and resistivity RT and corresponding historical lithofacies type information; Each historical lithofacies category in the input data file is mapped to a unique integer value, and the numerical features of all mapping results are standardized so that the mean of all numerical features is 0 and the variance is 1 to obtain the training data set.

[0012] Optionally, the method further includes balancing the sample quantities of historical well logging curve data and corresponding historical lithofacies type information in the training data set by using a synthetic minority class oversampling technique.

[0013] The above-mentioned lithofacies identification method for marine-continental transitional shale provided by the embodiment of the present invention has the following beneficial effects compared with the prior art: Compared with a single model, the present invention uses the lithofacies prediction results of each trained base learner as a new feature to form a new feature space. The new feature space is used to train the meta-classifier, and then the weights of the prediction outputs of different base learners in the fusion process are adjusted through the logistic regression model LogisticRegression. This can better capture the lithology of different types of marine-continental transitional shale and improve the robustness and stability of lithofacies identification and classification.

[0014] In addition, using at least two different types of machine learning models as base learners can fully mine and utilize the shale lithofacies information revealed by multidimensional logging data, capture the complex lithology of marine-continental transitional shale, and improve the accuracy of lithofacies identification. Using the logistic regression model LogisticRegression as a meta-classifier can handle multi-classification problems well and effectively capture the relationships between base learners. The output layer of the logistic regression can map the logging data to multiple categories and output the probability of each category, thereby achieving accurate classification of complex lithology. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1A schematic flow chart of a method for identifying lithofacies of marine-continental transitional shale provided in one embodiment; Figure 2 An algorithm flow chart of a lithofacies identification method for marine-continental transitional shale provided in one embodiment; Figure 3 Schematic diagram of an ROC curve for a lithofacies identification method of marine-continental transitional shale provided in one embodiment; Figure 4 A confusion matrix diagram of a lithofacies identification method for marine-continental transitional shale provided in one embodiment. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0017] In one embodiment, a method for identifying the lithofacies of marine-continental transitional shale is provided, such as Figure 1 As shown, the core steps of this method include: The historical well logging curve data and corresponding historical lithofacies type information of marine-continental transitional shale were obtained as the training data set.

[0018] At least two different types of machine learning models are used as base learners, and the logistic regression model LogisticRegression is used as a meta-classifier to build a stacking model.

[0019] The at least two different types of machine learning models include, but are not limited to, at least two of the following: an XGBoost model, a RandomForest model, a CatBoost model, and a LightGBM model. For example, the number of models may be two, three, or more. By adjusting the number and type of base learners, the algorithm can adapt to different types of lithologic classification tasks, whether simple binary classification problems or complex multi-classification problems. It can handle input data of various sizes and be expanded based on task requirements, ultimately achieving the goal of extending lithologic facies classification to the entire region.

[0020] The training dataset is input into the stacking model, and each base learner is trained using the training dataset to obtain multiple trained base learners. Each trained base learner is then used to predict the training dataset to obtain lithofacies predictions for the transitional shale. The lithofacies predictions of each trained base learner are used as new features to form a new feature space.

[0021] The meta-classifier is trained through the new feature space to obtain a trained meta-classifier to obtain a trained stacking model.

[0022] In the lithofacies identification process of the transitional shale, the well logging curve data of the transitional shale to be identified is input into the trained stacking model to obtain the lithofacies type identification result.

[0023] like Figure 2 As shown, a specific embodiment is provided by taking the XGBoost model, the Random Forest model, the CatBoost model, and the LightGBM model as examples: The historical well logging curve data and corresponding historical lithofacies type information of marine-continental transitional shale were obtained as the training data set.

[0024] The XGBoost model, Random Forest model, CatBoost model, and LightGBM model are used as base learners, and the logistic regression model LogisticRegression is used as the meta-classifier to build a stacking model.

[0025] The training data set is input into the stacking model, and each base learner is trained using the training data set to obtain multiple trained base learners. Each trained base learner is used to predict the training data set to obtain the lithofacies prediction results of the marine-continental transitional shale. The lithofacies prediction results of each trained base learner are used as new features to form a new feature space.

[0026] The meta-classifier is trained through the new feature space, and the weights of the lithofacies prediction results of different trained base learners in the fusion process are adjusted through the meta-classifier to obtain the trained meta-classifier to obtain the trained stacking model.

[0027] In the lithofacies identification process of the transitional shale, the well logging curve data of the transitional shale to be identified is input into the trained stacking model to obtain the lithofacies type identification result.

[0028] The specific implementation is: Step 1: Obtain an input data file containing historical well log data, including gamma ray (GR), acoustic travel time (AC), neutron porosity (CNL), density (DEN), and resistivity (RT), along with the corresponding historical lithofacies type information. Map each historical lithofacies type in the input data file to a unique integer value. Standardize all numerical features in the mapping results so that each feature has a mean of 0 and a variance of 1, thus obtaining the training dataset.

[0029] The Pandas library in the Python programming language is used to read the input data file in CSV format and load the data into memory. Next, a label encoder is used to encode the target variable (lithofacies type) into a numerical form to facilitate model processing. Label encoding maps each category to a unique integer value. A standard scaler is then used to normalize all numerical features to a mean of 0 and a variance of 1. This standardization process ensures that different features are on a similar scale, preventing certain features with larger numerical ranges from dominating the model training process.

[0030] Step 2: Use the Synthetic Minority Oversampling Technique (SMOTE) to balance the data and increase the number of minority class samples. SMOTE synthesizes new minority class samples to make the number of samples in each class in the dataset more balanced, thereby improving the model's ability to identify minority classes.

[0031] Step 3: Use the training set and test set split method to split the dataset into 80% for training and 20% for testing. The training set is used to train the model, while the test set is used to verify the model's performance and ensure that the model can perform well on unseen data.

[0032] Step 4: Optimize the hyperparameters of the underlying machine learning models (eXtreme GradientBoosting, XGBoost; Random Forest; CatBoost; and Light GBM). Hyperparameters are parameters that need to be manually set before training the model, rather than automatically adjusted by the learning algorithm. Tuning hyperparameters is critical to improving the performance of machine learning models because appropriately selecting hyperparameters can make the model more accurate and stable and reduce the risk of overfitting.

[0033] The XGBoost model is used as the base learner, specifically including: Determine the search range for the XGBoost model's hyperparameters. These include the learning rate, maximum tree depth, number of weak classifiers, sampling ratio, and the ratio of features used in each tree. During the grid search process, determine the values ​​for each hyperparameter in the XGBoost model, creating different hyperparameter combinations. Cross-validate each hyperparameter combination to maximize the accuracy of the XGBoost model and determine the optimal hyperparameter combination. The XGBoost model corresponding to the optimal hyperparameter combination is used as the hyperparameter-optimized XGBoost model, which is then used as the base learner.

[0034] The hyperparameter search range for the XGBoost model defines the learning rate, maximum tree depth, number of weak classifiers, sampling ratio, and the feature ratio used in each tree. A grid search is then performed to traverse the given hyperparameter combinations, performing cross-validation on each combination to maximize the model's accuracy and ultimately find the optimal hyperparameter combination for XGBoost.

[0035] Specifically: for the hyperparameter search range of the XGBoost model, the definitions include learning rate, maximum tree depth, number of weak classifiers, sampling ratio and feature ratio used by each tree; the learning rate parameter determines the step size of the model to update the parameters in each iteration, and the candidate value of the learning rate is set to [0.01, 0.02] to find a balance between stability and convergence speed; the maximum depth parameter controls the maximum depth of the XGBoost decision tree. During the search process, the maximum depth range is set to [2, 3] to explore the impact of different tree depths on model performance; the number of weak classifiers represents the number of decision trees in XGBoost, and the candidate value is set to [50, 100], aiming to find a value that can ensure the model The sample sampling ratio parameter is set to prevent overfitting, and its value range is [0.85, 0.9]. By partially sampling the training samples, each decision tree uses a different sample subset during training, which helps to improve the robustness and generalization ability of the model. The feature sampling ratio controls the feature ratio used when building the decision tree, and its value range is [0.85, 0.9]. This parameter can reduce the feature dimension and avoid overfitting. Through grid search, the hyperparameter combinations in the hyperparameter search space are traversed, and cross-validation is performed on each combination to maximize the accuracy of the XGBoost model and obtain the optimal hyperparameter combination of the XGBoost model.

[0036] The Random Forest model is used as the base learner, specifically including: Determine the search range for the hyperparameters of the Random Forest model. The hyperparameters of the Random Forest model include the number of decision trees, the maximum tree depth, the minimum number of split samples, the minimum number of samples per leaf node, and whether to perform bootstrap sampling. During the grid search process, calculate the impact of different combinations of hyperparameter values ​​on the performance of the Random Forest model and determine the hyperparameter combination with the optimal accuracy. The Random Forest model corresponding to the hyperparameter combination with the optimal accuracy is used as the hyperparameter-optimized Random Forest model, and the hyperparameter-optimized Random Forest model is used as the base learner.

[0037] Optimize the hyperparameters of the Random Forest model, including the number of decision trees, maximum tree depth, minimum number of split samples, minimum number of samples per leaf node, and whether to use bootstrap sampling. Grid search and cross-validation are used to find the optimal hyperparameter combination for the Random Forest model.

[0038] Specifically, the hyperparameters of the Random Forest model are optimized, including the number of decision trees, maximum tree depth, minimum number of split samples, minimum number of samples for leaf nodes, and bootstrap sampling. The number of decision trees is the number of weak classifiers. In a random forest, each decision tree contributes to the final prediction result, and the candidate values ​​are [10, 50, 100]. The maximum depth parameter controls the maximum depth of each decision tree in the random forest, and the candidate values ​​are set to [3, 5]. The minimum number of split samples represents the minimum number of samples required for node splitting, and the candidate values ​​are set to [5, 10]. By adjusting this parameter, the optimal sample splitting condition that can prevent overfitting and ensure model performance can be found. The candidate values ​​for the minimum number of samples for leaf nodes are set to [1, 2, 4]. During the search process, bootstrap is set to [True, False] for comparative experiments to determine the impact of using bootstrap sampling on model performance. During the grid search process, the impact of different value combinations of each hyperparameter on the performance of the RandomForest model is calculated to find the hyperparameter combination with the best accuracy on the training set.

[0039] The CatBoost model is used as the base learner, specifically including: Determine the search range for the hyperparameters of the CatBoost model. CatBoost model hyperparameters include the learning rate, tree depth, and number of iterations. During the grid search, traverse each hyperparameter combination of the CatBoost model in the hyperparameter grid. Evaluate the performance of the CatBoost model for each hyperparameter combination through cross-validation to determine the hyperparameter combination with the best performance. The CatBoost model corresponding to the optimal hyperparameter combination is used as the hyperparameter-optimized CatBoost model, which is then used as the base learner.

[0040] Hyperparameter optimization for the CatBoost model includes the learning rate, tree depth, and number of iterations. Grid search and cross-validation are used to find the optimal hyperparameter combination for the CatBoost model.

[0041] Specifically, the CatBoost model hyperparameters were optimized, including the learning rate, tree depth, and number of iterations. The candidate values ​​for the learning rate were set to [0.01, 0.02, 0.05] to strike a balance between training speed and final performance. The tree depth range was set to [3, 5, 7] to explore the impact of different tree depths on model performance. The iterations parameter represents the number of trees, or the number of iterations. Each tree continuously corrects the errors of the previous tree, which helps enhance the model's expressiveness. However, too many iterations may cause the model to overfit the training data, so the candidate values ​​for iterations were set to [50, 100]. The search process traversed each hyperparameter combination in the hyperparameter grid, trained the model for each combination, and evaluated its performance through cross-validation. Finally, the best-performing parameter combination was selected based on the evaluation metric (accuracy).

[0042] The LightGBM model is built using default parameters because the default parameters can already provide good performance. Further hyperparameter optimization will reduce the overall accuracy.

[0043] Step 5: Create a stacked model. In the first layer, a LightGBM model, an XGBoost model, a Random Forest model, and a CatBoost model built with default parameters are used as base learners. Each base learner is trained separately and used to make predictions on the validation subset. The prediction results of each base learner on the validation subset are recorded. For each base learner, its prediction output on the validation subset is extracted as a new feature, thus forming a new feature space.

[0044] In the second layer of the stacked model, the logistic regression model, LogisticRegression, is selected as the meta-classifier. This meta-classifier uses the new feature space output by the first-layer base learners as input. The target variable remains the label of the original data. During training, the logistic regression model is trained using the new feature space, learning how to combine the predictions of the different base learners and the weighted relationships between them, thereby improving the overall model performance.

[0045] Step 6: Use the trained stacked model to predict the test set and evaluate model performance. Use multiple metrics, such as accuracy, precision, recall, and F1 score, to comprehensively assess model performance. Receiver Operating Characteristic (ROC) curves and confusion matrices are plotted to further evaluate the model's classification performance.

[0046] Specifically, accuracy measures the proportion of samples correctly predicted by the model out of the total number of samples. This is calculated by using the trained stacked model to make predictions on the test set and comparing the predictions with the true labels of the test set. Precision is the proportion of samples predicted as positive that are actually positive. The calculation uses the "average" parameter, which first calculates the precision for each class separately. These are then weighted and summed based on the relative number of samples in each class in the test set to obtain the final weighted precision. Recall measures the proportion of samples correctly predicted as positive out of all actual positive samples. Setting "average" to "weighted" takes a weighted average of the recall for each class. The F1 score is the harmonic mean of the precision and recall. When the "average" parameter is set to "weighted," the F1 score is calculated by first calculating the precision and recall for each class. The weighted average is then weighted based on the number of samples in each class to obtain the weighted precision and recall. Finally, the weighted F1 score is calculated.

[0047] In one embodiment, a table of evaluation indicators of the present invention on the core well data of a certain study area is provided (as shown in Table 1), and a ROC curve of the present invention on the core well data of a certain study area is provided (as shown in Table 2). Figure 3 As shown, Figure 3 With the false positive rate as the horizontal axis and the true positive rate as the vertical axis), the confusion matrix of the core well data in a certain study area (such as Figure 4 shown).

[0048] Table 1 Evaluation metrics As shown in Table 1, the stacking model of the present invention achieved an accuracy of 0.937, a recall of 0.939, a precision of 0.931, and an F1 score of 0.927 on a training set of data from 10 cored wells in a certain study area. The stacking model of the present invention achieved an accuracy of 0.927, a recall of 0.922, a precision of 0.929, and an F1 score of 0.919 on a test set of data from 10 cored wells in a certain study area.

[0049] Depend on Figure 3 It can be obtained that the AUC (Area Under Curve, AUC, the area enclosed by the ROC curve and the horizontal axis) of the cored well A1 is 0.992, the AUC of the de-core well A2 is 0.998, the AUC of the de-core well A3 is 0.997, the AUC of the de-core well A4 is 0.997, the AUC of the de-core well B1 is 1.000, the AUC of the de-core well B2 is 1.000, the AUC of the de-core well C1 is 0.998, the AUC of the de-core well D is 0.997, and the AUC of the de-core well E is 0.996.

[0050] Furthermore, the input of the present invention can be a combination of multiple logging curves, and the comprehensive effect of multiple logging parameters will be automatically processed by the model, thereby being suitable for high-dimensional space classification.

[0051] Furthermore, the present invention balances the data through the synthetic minority oversampling technology SMOTE, increases the number of minority class samples, solves the class imbalance problem, and makes the number of samples in each category in the data set more balanced. This enables the model to better identify and classify minority class samples during training, thereby improving the overall classification performance.

[0052] Furthermore, the present invention can also combine multiple optimization algorithms to accelerate training and improve classification performance, such as using random search or Bayesian optimization for hyperparameter optimization, helping the model to converge quickly, find the global optimal solution, and find the optimal weight combination in the complex loss function space, thereby improving classification performance.

[0053] Furthermore, the present invention uses logistic regression as a meta-classifier, which can handle multi-classification problems well and effectively capture the relationships between base learners. The output layer of logistic regression can map the logging data to multiple categories and output the probability of each category, thereby achieving accurate classification of complex lithology.

[0054] The present invention has the following advantages: (1) This paper solves the class imbalance problem by using SMOTE to balance the data. This enables the model to better identify and classify minority class samples during training, improving the overall classification performance.

[0055] (2) This paper significantly improves the accuracy of shale lithofacies classification by constructing a stacking model and combining the advantages of multiple machine learning algorithms. Compared with a single model, the stacking model can better capture the complexity of the data and improve the robustness and stability of the classification.

[0056] (3) This paper optimizes the hyperparameters of basic models such as XGBoost, Random Forest, LightGBM, and CatBoost through grid search. This enables the model to find the optimal hyperparameter combination, further improving classification performance.

[0057] (4) The present invention can perform intelligent quantitative identification of different shale lithofacies based on well logging curves, and the shale lithofacies prediction accuracy in the study area reaches 92%.

[0058] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A lithofacies identification method for marine-continental transitional shale, characterized by: include: Obtain historical well logging data and corresponding historical lithofacies type information of marine-continental transitional shale as a training dataset; Use at least two different types of machine learning models as base learners and the logistic regression model LogisticRegression as a meta-classifier to build a stacked model; The training data set is input into the stacking model, and each base learner is trained using the training data set to obtain multiple trained base learners. Each trained base learner is used to predict the training data set to obtain the lithofacies prediction results of the marine-continental transitional shale. The lithofacies prediction results of each trained base learner are used as new features to form a new feature space. The meta-classifier is trained using the new feature space to obtain a trained meta-classifier, thereby obtaining a trained stacking model. In the lithofacies identification process of the transitional shale, the well logging curve data of the transitional shale to be identified is input into the trained stacking model to obtain the lithofacies type identification result.

2. The method for identifying facies of marine-continental transitional shale according to claim 1, wherein: The at least two different types of machine learning models include: an XGBoost model, a Random Forest model, a CatBoost model, and a LightGBM model.

3. The method for identifying facies of marine-continental transitional shale according to claim 2, characterized in that: The XGBoost model is used as a base learner, specifically including: Determine the search range of the hyperparameters of the XGBoost model; the hyperparameters of the XGBoost model include: learning rate, maximum tree depth, number of weak classifiers, sampling ratio, and feature ratio used by each tree; During the grid search process, the value of each hyperparameter of the XGBoost model is determined to form different hyperparameter combinations; Perform cross-validation on each hyperparameter combination to maximize the accuracy of the XGBoost model and obtain the optimal hyperparameter combination; The XGBoost model corresponding to the optimal hyperparameter combination is used as the XGBoost model after hyperparameter optimization, and the XGBoost model after hyperparameter optimization is used as the base learner.

4. The method for identifying facies of marine-continental transitional shale according to claim 2, wherein: The Random Forest model is used as a base learner, specifically including: Determine the search range of hyperparameters of the Random Forest model; the hyperparameters of the Random Forest model include: the number of decision trees, the maximum tree depth, the minimum number of split samples, the minimum number of samples for leaf nodes, and whether to perform self-service sampling; During the grid search process, the effects of different combinations of hyperparameters of the Random Forest model on the Random Forest model are calculated to determine the hyperparameter combination with the best accuracy. The Random Forest model corresponding to the hyperparameter combination with the best accuracy is used as the RandomForest model after hyperparameter optimization, and the Random Forest model after hyperparameter optimization is used as the base learner.

5. The method for identifying facies of marine-continental transitional shale according to claim 2, wherein: The CatBoost model is used as a base learner, specifically including: Determine a search range for hyperparameters of a CatBoost model; the hyperparameters of the CatBoost model include: learning rate, tree depth, and number of iterations; During the grid search process, each hyperparameter combination of the CatBoost model in the hyperparameter grid is traversed; Evaluate the performance of the CatBoost model under each hyperparameter combination through cross-validation to determine the hyperparameter combination with the best performance; The CatBoost model corresponding to the best hyperparameter combination is used as the CatBoost model after hyperparameter optimization, and the CatBoost model after hyperparameter optimization is used as the base learner.

6. The method for identifying lithologic features of marine-continental transitional shale according to claim 1, wherein: The acquisition of historical well logging data and corresponding historical lithofacies type information of marine-continental transitional shale as a training data set specifically includes: Obtain input data files including historical well logging data of gamma ray GR, acoustic travel time AC, neutron porosity CNL, density DEN and resistivity RT and corresponding historical lithofacies type information; Each historical lithofacies category in the input data file is mapped to a unique integer value, and the numerical features of all mapping results are standardized so that the mean of all numerical features is 0 and the variance is 1 to obtain the training data set.

7. The method for identifying facies of marine-continental transitional shale according to claim 1, wherein: The method also includes balancing the number of samples of historical well logging curve data and corresponding historical lithofacies type information in the training data set by using a synthetic minority class oversampling technology.

Images