Low-permeability oil and gas reservoir classification method based on interpretable machine learning

Abstract

The invention discloses a low-permeability oil and gas reservoir classification method based on interpretable machine learning, and the method comprises the following steps: obtaining a core sample of a target block, and obtaining core parameters; calculating a Pearson's correlation coefficient among all the core parameters, and if the correlation coefficient between two core parameters is greater than a first threshold value, randomly rejecting one core parameter to obtain an initial data set; based on the initial data set, generating a plurality of resampling data sets by utilizing a bootstrap technology; based on a random forest model, training a decision tree for any resampling data set, and selecting an optimal hyper-parameter for the model by using a grid search method; and screening out a plurality of important features based on the Shapley addition interpretation and the average impurity reduction, and interpreting different core categories of the target block based on a principal component analysis method. According to the method, an interpretable machine learning method is introduced, different sedimentary facies in the low-permeability reservoir are classified, and corresponding data support can be provided for subsequent oil and gas development of a block.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning and geology, and in particular to a low-permeability oil and gas reservoir classification method based on interpretable machine learning. Background Art

[0002] Low-permeability reservoirs have become the target of oil and gas exploration in lacustrine basins. However, the complex pore-throat structure and the influence of diagenesis may hinder our understanding of reservoir quality. In addition, it is challenging to identify different types of low-permeability reservoirs, such as beach-bar and turbidite deposits, which are common in lacustrine basins, due to the low permeability and porosity. It is difficult to distinguish these types using traditional statistical analysis of mercury injection pore simulation. Summary of the invention

[0003] To solve at least one of the above problems, the present invention proposes a low permeability oil and gas reservoir classification method based on interpretable machine learning.

[0004] The technical solution of the present invention is: a low permeability oil and gas reservoir classification method based on interpretable machine learning, comprising the following steps:

[0005] S1. Obtain core samples of at least two core categories in a target block, and obtain core parameters of the core samples;

[0006] S2, calculating the Pearson correlation coefficient between all core parameters, if the correlation coefficient between two core parameters is greater than a first threshold, randomly eliminating one to obtain an initial data set;

[0007] S3, based on the initial data set, generate a resampled data set using the bootstrap technique until the sample size in the resampled data set is the same as the sample size in the initial data set, and repeat the above operation until multiple resampled data sets are generated;

[0008] S4. Based on the random forest model, a decision tree is trained for any resampled data set. Based on multiple resampled data sets, the grid search method is used to select the best hyperparameters for the model, and finally the trained model is obtained;

[0009] S5. Based on the Shapley additive interpretation and the average reduction of impurities, the features in the random forest model are interpreted to evaluate the importance of each feature in the core parameters, screen out multiple important features, and combine the categories of core samples and the differences in important features of each category. Based on the principal component analysis method, different core categories in the target block are interpreted.

[0010] Beneficial effects:

[0011] The present invention introduces an interpretable machine learning method, which combines the random forest classifier with the SHAP and MDI feature importance indicators to classify different sedimentary phases in low permeability reservoirs, and can provide corresponding data support for subsequent oil and gas development in this block. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a SHAP summary diagram and an MDI summary diagram of an embodiment of the present invention;

[0013] Figure 2 This is the PCA analysis diagram of this embodiment. DETAILED DESCRIPTION

[0014] The specific implementation modes of the present invention will be described clearly and completely below with reference to examples and drawings. Obviously, the described examples are only some embodiments of the present invention, rather than all embodiments.

[0015] Example 1

[0016] A low permeability oil and gas reservoir classification method based on interpretable machine learning includes the following steps:

[0017] S1. Obtain core samples of at least two core categories in a target block, and obtain core parameters of the core samples;

[0018] In this embodiment, the core category refers to the sedimentary phase category. At the same time, this embodiment is used to identify the low permeability of different sedimentary phase categories. Therefore, the selected core samples are usually sedimentary phases that are difficult to distinguish, such as turbidite and beach-bar rock.

[0019] In this embodiment, the core parameters are porosity, permeability, average pore throat radius, displacement pressure, maximum pore throat radius, mercury removal efficiency, mercury exclusion volume, structure coefficient, characteristic coefficient, root mean square, pore throat sorting coefficient, relative sorting coefficient, kurtosis, skewness, specific surface area, surface area-volume ratio, macroporosity, micropore permeability, and pore volume, and these parameters are obtained by mercury injection experiment.

[0020] In this embodiment, a total of 194 core samples are obtained. For each core sample, its core category and core parameters need to be obtained. These core categories include turbidite and beach-bar rock.

[0021] S2, calculating the Pearson correlation coefficient between all core parameters, if the correlation coefficient between two core parameters is greater than a first threshold, randomly eliminating one to obtain an initial data set;

[0022] The Pearson correlation coefficient is a data that measures the correlation between two parameters. Its value is distributed between -1 and 1. The closer the absolute value of the Pearson correlation coefficient is to 1, the higher the correlation between the two parameters. The closer the absolute value is to 0, the lower the correlation between the two parameters.

[0023] Since the random forest model is used in this embodiment, the increase in the number of highly correlated parameters does not enhance the performance of the model. Therefore, it is necessary to randomly delete the parameters with higher correlation. For the first threshold, it can be set to 0.8-0.9, and those skilled in the art can set it according to actual conditions. In this embodiment, the first threshold is selected as 0.9. Finally, 6 parameters were deleted, and the remaining parameters were: porosity, permeability, specific surface area, micropore porosity, structural coefficient, characteristic coefficient, displacement pressure, mercury extraction, kurtosis, efficiency of mercury extraction, pore throat sorting coefficient, average pore throat radius, and skewness.

[0024] The data set after removing the highly correlated parameters is used as the initial data set of this embodiment.

[0025] S3, based on the initial data set, generate a resampled data set using the bootstrap technique until the sample size in the resampled data set is the same as the sample size in the initial data set, and repeat the above operation until multiple resampled data sets are generated;

[0026] In this step, firstly, random sampling with replacement is performed based on the initial data set to generate a resampled data set with the same sample size as the initial data set. Then, the above operation is repeated to generate multiple resampled data sets. The number of data sets finally generated can be set according to the actual situation, such as 1000 data sets, etc.

[0027] S4. Based on the random forest model, a decision tree is trained for any resampled data set. Based on multiple resampled data sets, the grid search method is used to select the best hyperparameters for the model, and finally the trained model is obtained;

[0028] After generating multiple resampled data sets, a decision tree is trained for each resampled data set using the random forest model. Since in the random forest model, each decision tree needs to select a suitable feature selection metric to maximize the similarity between classes. In this embodiment, the grid search method is used to select the best hyperparameter value for the model. The choice of hyperparameters will have a significant impact on the overall performance of the model; therefore, the grid search method will exhaustively explore all combinations of hyperparameters to achieve the best performance. The specific steps of the grid search method are as follows:

[0029] Define the hyperparameter space and set a value range for each hyperparameter. The values ​​of these hyperparameters constitute a search space. For hyperparameters, they usually include the number of classifiers and the depth of decision trees.

[0030] The candidate values ​​of each hyperparameter are combined to form a grid. Then, based on multiple reused datasets of S3, the model is trained and evaluated using five-fold cross validation, and the better model is selected using accuracy and area under the curve.

[0031] S5. Based on the Shapley additive interpretation and the average reduction of impurities, the features in the random forest model are interpreted to evaluate the importance of each feature in the core parameters, screen out multiple important features, and combine the categories of core samples and the differences in important features of each category. Based on the principal component analysis method, different core categories in the target block are interpreted.

[0032] Get the SHAP value and MDI value for each feature, and build a SHAP summary graph and an MDI summary graph; in the SHAP algorithm, the importance of any feature is the weighted sum of the contribution of the feature to the model output in all possible feature combinations. By quantifying the importance of each feature, the SHAP value can identify the features that drive the model prediction. SHAP values ​​help explain the contribution of features in complex ensemble models such as random forests. The most influential features have the highest SHAP values, and the least influential features have the lowest SHAP values. Similarly, MDI is determined by the depth of certain nodes in the RFC decision tree. Features at the top of the decision tree contribute more to the final decision than features at the bottom. Relative importance is measured using this score. MDI combines this fraction with the reduction in entropy impurity to calculate the importance of the feature; therefore, the most influential features have the highest MDI.

[0033] According to the SHAP summary graph and the MDI summary graph, the important features that affect the model decision are screened out; the reason for establishing the summary graph is that the importance of the features in the model can be observed more intuitively through the summary graph. Figure 1 As shown. Figure 1 It can be seen that the importance of the four characteristics of specific surface area, permeability, porosity and microporosity is much higher than the importance of the other parameters, and the SHAP summary diagram and the MDI summary diagram can verify each other. Through further comparison, it was found that the ranking of MDI values ​​was slightly different from that of SHAP values: the MDI value of specific surface area was higher than the MDI value of permeability, and the SHAP value of permeability was higher than the SHAP value of specific surface area.

[0034] After screening out the important features that affect the model decision through the SHAP summary diagram and the MDI summary diagram, the principal component analysis method is also used in this embodiment to interpret these important features, so as to distinguish the characteristics of the turbidite and beach-bar rock in this block in terms of mercury injection parameters, and provide corresponding data support for the subsequent development of this block.

[0035] When using the principal component analysis method for interpretation, the four features of surface area, permeability, porosity and microporosity are first reduced in dimension. Then, based on the characteristics of the reduced features, two principal components PC1 and PC2 are set. Then, the general operation of the principal component analysis method is used to analyze them. The final results are as follows: Figure 2 As shown. Figure 2 It can be seen that the two principal components explain 47.19% and 37.29% of the variance in the data set, respectively, indicating that these two features can classify beach-bar rocks and turbidites. Among them, beach-bar sandstone has higher porosity, high permeability, and low specific surface area, while turbidites have lower porosity, low permeability, and high specific surface area.

[0036] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A low permeability oil and gas reservoir classification method based on interpretable machine learning, characterized in that: The following steps are involved: S1. Obtain core samples of at least two core categories in a target block, and obtain core parameters of the core samples; S2, calculating the Pearson correlation coefficient between all core parameters, if the correlation coefficient between two core parameters is greater than a first threshold, randomly eliminating one to obtain an initial data set; S3, based on the initial data set, generate a resampled data set using the bootstrap technique until the sample size in the resampled data set is the same as the sample size in the initial data set, and repeat the above operation until multiple resampled data sets are generated; S4. Based on the random forest model, a decision tree is trained for any resampled data set. Based on multiple resampled data sets, the grid search method is used to select the best hyperparameters for the model, and finally the trained model is obtained; S5. Based on the Shapley additive interpretation and the average reduction of impurities, the features in the random forest model are interpreted to evaluate the importance of each feature in the core parameters, screen out multiple important features, and combine the categories of core samples and the differences in important features of each category. Based on the principal component analysis method, different core categories in the target block are interpreted.

2. The method according to claim 1, characterized in that The core parameters include porosity, permeability, specific surface area, microporosity, structural coefficient, characteristic coefficient, displacement pressure, mercury extraction amount, kurtosis, efficiency of mercury extraction, pore throat sorting coefficient, average pore throat radius, and skewness.

3. The method according to claim 1, characterized in that The category of the core sample refers to its sedimentary facies category.

4. The method according to claim 1, characterized in that: The first threshold is 0.8-0.

9.

5. The method according to claim 1, characterized in that: In S4, The following steps are included: For each resampled data set of S3, a decision tree is trained; For decision trees, define a hyperparameter space and set a value range for each hyperparameter. The values ​​of these hyperparameters constitute a search space. The candidate values ​​of each hyperparameter are combined to form a grid. Based on multiple resampled data sets of S3, the model is trained and evaluated using five-fold cross validation. The better model is then selected using accuracy and area under the curve.

6. The method according to claim 1, characterized in that S5 includes the following sub-steps: Get the SHAP value and MDI value of each feature, and create SHAP summary graph and MDI summary graph; According to the SHAP summary diagram and MDI summary diagram, the important features that affect the model decision are screened out; Based on principal component analysis, the important features were interpreted and identified according to their differences in variance in the data sets of different core categories.

Images