Double-layer learner integrated model for predicting porosity of mixed shale

Abstract

The invention discloses a double-layer learner integrated model for predicting the porosity of mixed shale, and relates to the technical field of mixed shale porosity evaluation. According to the technical scheme, the method comprises the following steps that S1, logging data with high correlation with the actually-measured porosity are selected; s2, constructing an optimal learning device; and S3, constructing a double-layer learner integration model. According to the method, the logging data with high correlation with the porosity are preferably selected through a principal component analysis method, so that noise pollution caused by input of a large amount of logging data is avoided; the optimal hyper-parameter value of the machine learning model is searched by using a particle swarm optimization (PSO) algorithm, so that the learning ability of a base learner is improved to the greatest extent; the base learner is used as the first layer learner of the integrated model, and the meta learner of the second layer is combined, so that the integrated model of the double-layer learner is constructed, the purposes of improving the model prediction precision and evaluating the vertical heterogeneity of the pores of the mixed shale reservoir are achieved, and the purpose of guiding unconventional oil-gas exploration and development of the mixed shale is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mixed shale porosity evaluation, and particularly relates to a double-layer learner integrated model for predicting mixed shale porosity. BACKGROUND

[0002] With the continuous deepening and development of oil and gas exploration and development, unconventional oil and gas resources represented by shale oil have shown great potential under the existing economic and technical conditions, and the successful exploitation of shale oil in mixed shale (high content of siliceous, calcareous and clay minerals) has further put forward higher requirements for physical property evaluation. However, the physical property of mixed shale reservoir changes frequently and is generally poor, and shale oil mainly accumulates in micro-nano scale pores, so porosity becomes a key parameter for evaluating the shale oil potential of mixed shale.

[0003] However, for conventional logging interpretation methods (such as inversion method, empirical formula method and multiple regression method, etc.), due to the difficulty in selecting key parameters (such as baseline), the prediction accuracy of porosity is often low and the generalization ability is poor. Machine learning methods (such as random forest, support vector machine, extreme learning machine, K nearest neighbor algorithm, fuzzy logic regression and Bayesian regression, etc.) can directly start from the characteristics of data itself, but there are still great difficulties in continuously and accurately evaluating porosity from the field scale. The main reasons are that the "noise" interference carried in a large amount of logging data, the influence of hyperparameters in machine learning model on learning ability, and the inherent limitations of single machine learning algorithm, all bring great challenges to fine porosity evaluation, and seriously limit the pace of unconventional oil and gas exploration and development of mixed shale. Therefore, it is necessary to find a suitable machine learning model integration framework and apply it to porosity evaluation, reveal the vertical physical property heterogeneity of mixed shale reservoir, and guide the selection of shale oil "sweet spot" and its exploration and development. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a double-layer learner integrated model for predicting the porosity of mixed shale. The principal component analysis method is used to select logging data with strong correlation with porosity, so as to avoid noise pollution caused by a large amount of logging data input. The particle swarm optimization algorithm (PSO) is used to find the optimal hyperparameter value of the machine learning model, so as to maximize the learning ability of the base learner. The base learner is used as the first layer learner of the integrated model, and the meta-learner of the second layer is combined, so as to construct the integrated model of the double-layer learner, so as to improve the prediction accuracy of the model, evaluate the vertical heterogeneity of the porosity of the mixed shale reservoir, and guide the unconventional oil and gas exploration and development of the mixed shale.

[0005] The technical scheme of the present application is as follows:

[0006] The double-layer learner integrated model for predicting the porosity of mixed accumulation shale comprises the following steps:

[0007] S1: selecting well logging data with strong correlation with measured porosity: performing principal component analysis on the well logging data, determining the optimal input feature dimension based on the cumulative variance contribution rate of the well logging data, and then selecting the top several well logging data with the strongest correlation with the measured porosity, wherein the number of selected well logging data is the same as the input feature dimension, thereby providing effective input data for the subsequent model training and evaluation;

[0008] S2: constructing an optimal learner: dividing the well logging data and the measured porosity selected in step S1 into a training set and a test set, wherein the data amount in the training set is greater than that in the test set, and the specific data amount ratio can be divided according to the actual situation; using K nearest neighbor algorithm (KNN), random forest algorithm (RF), extreme gradient boosting algorithm (XGBoost), adaptive boosting algorithm (AdaBoost), lightweight gradient boosting algorithm (LightGBM) and category boosting algorithm (CatBoost) as the first layer learner, performing learner training on the training set, obtaining the optimal value of the hyperparameters of each learner, and verifying the training result of the learner on the test set;

[0009] S3: constructing a double-layer learner integrated model: combining two or more of the above-mentioned learners, wherein the optimal value of the hyperparameters of the learners is obtained in step S2; stacking the prediction accuracy of each learner combination on the total data set to form a new data set; taking the new data set as the input feature and the measured porosity as the output target, performing model training on the second layer meta-learner, and finally constructing a double-layer learner integrated model.

[0010] In step S1, when determining the optimal input feature dimension, the cumulative variance contribution rate of the principal component analysis result is used as the judgment standard instead of the traditional method of selecting the eigenvalue and scatter plot as the judgment standard, which can intuitively and quantitatively reflect the amount of information contained in the principal component of the original data. When determining the optimal input feature dimension according to the principal component analysis result, the input feature quantity corresponding to the inflection point of the curve in the graph can be selected as the optimal input feature dimension, at this time the input feature has retained most of the information of the original data, and the input feature quantity is not too much, thereby being able to extract the main feature information while reducing the data dimension.

[0011] Preferably, in step S1, the correlation between the well logging data and the measured porosity is evaluated by the following method: drawing a correlation curve of the measured porosity with respect to the well logging data to obtain the correlation coefficient of the curve in the graph, and the greater the correlation coefficient, the stronger the correlation between the well logging data and the measured porosity.

[0012] Preferably, in step S1, when performing principal component analysis on the logging data, the logging data is normalized and mapped to the [0, 1] interval to eliminate the influence of outliers (i.e., excessively large or small values ​​in the logging data) on the principal component analysis.

[0013] Preferably, in step S2, the hyperparameters are evaluated as having reached their optimal values ​​by the following method: the hyperparameter value corresponding to the maximum prediction accuracy of the training set is the optimal hyperparameter value.

[0014] Preferably, in step S3, a linear regression function, KNN, RF, XGBoost, AdaBoost, LightGBM, or CatBoost is used as the meta-learner.

[0015] Preferably, in step S3, a linear regression function is used as the meta-learner.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. This invention uses principal component analysis to select well logging data that are strongly correlated with porosity, thus avoiding noise pollution caused by a large amount of well logging data input; it uses particle swarm optimization (PSO) to find the optimal hyperparameter values ​​of the machine learning model to maximize the learning ability of the base learner; it uses the base learner as the first layer learner of the integrated model, combined with the second layer meta learner, to construct a two-layer integrated model, thereby improving the model's prediction accuracy and evaluating the vertical heterogeneity of porosity in mixed shale reservoirs, thus guiding the exploration and development of unconventional oil and gas in mixed shale.

[0018] 2. This invention considers the objective reality that current well logging data contains a large amount of noise, making it difficult for machine learning models to achieve ideal prediction accuracy. Therefore, based on principal component analysis, it selects key well logging data to minimize data contamination while maximizing the inclusion of original data information, providing ideal input data for accurate subsequent porosity evaluation. Simultaneously, this invention fully considers the impact of hyperparameters on the learning ability of machine learning models, proposing a method based on particle swarm optimization (PSO) to find suitable hyperparameter values, rather than traditional manual selection and grid search, to maximize the model's learning ability and provide an ideal base learner for subsequent ensemble model construction. The proposed two-layer ensemble model is easy to construct, the entire scheme is simple to operate, and highly feasible, providing a new method for the fine evaluation of porosity in mixed shale reservoirs and serving as a demonstration for the exploration and development of unconventional oil and gas.

[0019] 3. This invention solves the problem of low accuracy in porosity evaluation based on well logging data, which is currently quite challenging. Compared with conventional well logging analysis and early single machine learning methods, it has higher accuracy and reliability, and provides important technical support for the exploration and development of mixed sedimentary shale oil and gas. Attached Figure Description

[0020] Figure 1 This is the principal component analysis result of the well logging data in Embodiment 1 of the present invention.

[0021] Figure 2 This is a correlation curve between well logging data and porosity in Embodiment 1 of the present invention.

[0022] Figure 3 This is the porosity prediction result of the two-layer learner ensemble model constructed in Embodiment 1 of the present invention, which uses LightGBM-CatBoost as the first-layer base learner and a linear regression function as the second-layer meta-learner. Detailed Implementation

[0023] This invention takes the fine evaluation of porosity of the mixed sedimentary shale of the second member (E1f2) of the Funing Formation in the X1 well of the Gaoyou Depression in the Subei Basin as an example, and describes the specific implementation scheme and application effect of the invention in conjunction with the accompanying drawings and specific embodiments.

[0024] The two-layer learner ensemble model for predicting the porosity of mixed-sediment shale of the present invention includes the following steps:

[0025] S1 Selects logging data that are highly correlated with measured porosity: Principal component analysis is performed on the logging data, and the optimal input feature dimension is determined based on the cumulative variance contribution rate of the logging data; then, the top few logging data that are most strongly correlated with the measured porosity of mixed shale are selected, and the number of logging data selected is the same as the input feature dimension.

[0026] S2 Constructing the optimal learner: Divide the well logging data and measured porosity selected in step S1 into a training set and a test set; with measured porosity as the output target, use KNN, RF, XGBoost, AdaBoost, LightGBM and CatBoost as the first-layer learners respectively, train the learners on the training set to obtain the optimal hyperparameter values ​​of each learner, and verify the training results of the learners on the test set;

[0027] S3 Constructing a two-layer learner ensemble model: Combining two or more learners, wherein the hyperparameter values ​​of the learners are all the optimal hyperparameter values ​​obtained in step S2; stacking the prediction accuracy of each learner combination on the total dataset column-wise to form a new dataset; using the new dataset as input features and the measured porosity as the output target, training the second-layer meta-learner model, and finally constructing the two-layer learner ensemble model.

[0028] In step S1, the correlation between logging data and measured porosity is evaluated by the following method: a correlation curve is plotted between the measured porosity and the logging data, and the correlation coefficient of the curve in the figure is obtained. The larger the correlation coefficient, the stronger the correlation between the logging data and the measured porosity.

[0029] In step S1, when performing principal component analysis on the logging data, the logging data is normalized and mapped to the interval [0, 1] to eliminate the influence of outliers in the logging data on the principal component analysis.

[0030] In step S2, the following method is used to judge whether the hyperparameters have reached their optimal values: the hyperparameter values ​​corresponding to the maximum prediction accuracy of the training set are the optimal hyperparameter values.

[0031] In step S3, linear regression, KNN, RF, XGBoost, AdaBoost, LightGBM, and CatBoost are used as meta-learners. Preferably, linear regression is used as the meta-learner in step S3.

[0032] Example 1

[0033] To preserve the original porosity characteristics of the mixed-phase shale, the mixed-phase shale samples selected in this embodiment are mainly frozen mixed-phase shale core samples from different depths of E1f2 in Well X1 in the Gaoyou Depression of the Subei Basin. Studies show that the mineral composition of the shale is complex, and vertically it can be divided into six lithofacies: felsic argillaceous mixed shale, felsic gray dolomitic mixed shale, argillaceous felsic mixed shale, gray dolomitic argillaceous felsic shale, gray dolomitic shale, and felsic shale. The porosity development characteristics of different lithofacies vary, and coupled with the noise pollution inherent in the well logging data, traditional well logging interpretation methods and single machine learning models struggle to achieve ideal predictive results. Therefore, this embodiment proposes a two-layer model integration framework based on principal component analysis of well logging data and hyperparameter optimization of machine learning models, attempting to integrate multiple machine learning models to improve the model's learning ability and generalization capability.

[0034] This embodiment performs principal component analysis on well logging data, and the results are as follows: Figure 1As shown in the figure, it can be seen that as the number of input features increases, the variance contribution rate of each principal component shows a decreasing trend, indicating that the amount of information carried by the input features gradually decreases; when the number of input features is 6, the cumulative variance contribution rate of each principal component reaches 0.967, indicating that the input features have retained most of the information of the original data, thus determining that the optimal dimension of the input features is 6.

[0035] Subsequently, a correlation analysis was performed between well logging data and porosity, resulting in a correlation curve, such as... Figure 2 As shown. Because well logging data represents a comprehensive response to all minerals, organic matter, and pore fluids in the whole rock, and because the mineral composition of mixed shale is quite complex (siliceous, calcareous, and argillaceous minerals account for nearly 30%), the correlation between well logging data and its porosity is generally relatively low. Figure 2 It can be seen that, among the well logging data, GR logging data has the strongest positive correlation with porosity (R... 2 =0.354), followed by KTH logging data (R 2 =0.256); AC logging data and porosity show a negative correlation, with a correlation coefficient R. 2 The value was 0.217; K, Th, and U logging data reflect the potassium, thorium, and uranium contents of the formation, respectively, and show a weak correlation with porosity. 2 The values ​​are 0.181, 0.0843, and 0.107, respectively. Combined with... Figure 1 The principal component analysis results are displayed, and these 6 well logging curves are selected as input features to construct the integrated model.

[0036] Next, this embodiment constructs an optimal learner. The PSO algorithm is used to optimize the learner's hyperparameters to achieve high-precision prediction of porosity data. Specifically, the selected well logging data and porosity are divided into training and testing sets in a 4:1 ratio. With measured porosity as the output target, KNN, RF, XGBoost, AdaBoost, LightGBM, and CatBoost are used as the first-layer learners. The learners are trained on the training set to obtain the optimal hyperparameter values ​​for each learner. The training results are then validated on the testing set.

[0037] The optimized learner's hyperparameter values ​​and prediction accuracy are shown in Table 1:

[0038] Table 1. Comparison of hyperparameter optimization results and prediction accuracy of different learners

[0039]

[0040]

[0041] In Table 1, 'kd_tree' refers to using a KD-tree (K-Dimensional Tree) as the data structure to store and train data; 'sqrt' means that when building the decision tree, the number of features considered for each split is the square root of the total number of features; 'gbtree' refers to using a tree-based model as the base learner; 'square' refers to using squared error as the loss function in the boosting process; 'gbdt' refers to using Gradient Boosting Decision Tree (GBDT) as the boosting method; and 'RMSE' refers to using root mean square error as the loss function to evaluate the model's predictive performance.

[0042] The learner used in this invention is developed based on the Python language and utilizes the relevant algorithms and functions provided in the sk.learn 1.2.2 library. As shown in Table 1, due to their higher structural complexity and learning ability, the CatBoost and LightGBM models significantly outperform other models in prediction tasks.

[0043] Two or more of the above learners are combined, and the hyperparameter values ​​of the learners are all the optimal hyperparameter values ​​obtained from Table 1. The prediction accuracy of each learner combination on the total dataset is stacked column by column to form a new dataset. The new dataset is used as the input feature and the measured porosity is used as the output target to train the second-layer meta-learner, and finally a two-layer learner ensemble model is constructed.

[0044] Table 2 compares the performance of different two-layer heterogeneous ensemble models on prediction tasks, where the second-layer meta-learner of the two-layer heterogeneous ensemble model uses a linear regression function.

[0045] Table 2. Prediction accuracy R of ensemble models composed of different base learners and linear regression functions (meta-learners) 2

[0046]

[0047] In Table 2, XgbB: XGBoost; AdaB: AdaBoost; LGBM: LightGBM; CatB: CatBoost.

[0048] As shown in Tables 1 and 2, the ensemble of two-layer heterogeneous models affects the prediction accuracy. The performance of the first-layer base learner has a significant impact on the prediction performance of the ensemble model. As shown in Table 1, the LightGBM and CatBoost learners have higher prediction accuracy. Therefore, compared with other ensemble models, the ensemble model composed of the LGBM-CatB base learner (LightGBM and CatBoost) and the linear regression function meta-learner has the highest prediction accuracy R. 2(0.901); Furthermore, Table 2 also shows that the number of base learners does not necessarily improve the prediction performance of the ensemble model. For example, the prediction accuracy R of the KNN-XgbB-AdaB-LGBM-CatB ensemble model is low. 2 It is only 0.745.

[0049] Table 3 compares the performance impact of different meta-learners on the prediction task of the two-layer learner ensemble model:

[0050] Table 3. Prediction accuracy (R) of different two-layer learner ensemble models for predicting porosity in mixed shale. 2

[0051]

[0052] Taking KNN-XGBoost, RF-AdaBoost, and LightGBM-CatBoost models as examples, when the meta-learner of the second layer of the ensemble framework uses a linear regression function, the prediction accuracy R of the ensemble model... 2 The value is highest; however, when the meta-learner uses the relatively complex XGBoost model, the prediction accuracy R of the ensemble model is lower. 2 The value is the lowest. Based on the aforementioned experimental results, it can be concluded that the base learner in the first layer of the integrated framework should have strong learning and expressive capabilities, while the meta-learner in the second layer should have low model complexity. Overall, the prediction results of the two-layer model are significantly better than those of a single machine learning model, which may become a solution to the difficulty of porosity prediction in complex lithological reservoirs (such as mixed shale, sandstone, and conglomerate).

[0053] Figure 3 The figure shows the porosity prediction results of the ensemble model constructed in this embodiment, which uses LightGBM-CatBoost as the first-layer base learner and a linear regression function as the second-layer meta-learner, in the E1f2 section of well X1. As can be seen from the figure, the ensemble model can accurately predict the porosity change trend in different porosity ranges, and the error with the measured porosity is small. This indicates that the two-layer learner ensemble model has strong generalization ability and robustness, and can adapt to the porosity prediction task under mixed sedimentary shale conditions.

Claims

1. A dual-layer learner ensemble model for predicting porosity of a mixed accumulation shale, characterized in that, The method comprises the following steps: S1: selecting well logging data with strong correlation with measured porosity: performing principal component analysis on the well logging data, and determining the optimal input feature dimension based on the cumulative variance contribution rate of the well logging data; Then select the top several well logging data with the strongest correlation with the measured porosity of the mixed shale, and the number of selected well logging data is the same as the input feature dimension; S2: constructing an optimal learner: dividing the well logging data and the measured porosity selected in step S1 into a training set and a test set; Using the measured porosity as the output target, using K nearest neighbor algorithm, random forest algorithm, extreme gradient boosting algorithm, adaptive boosting algorithm, light gradient boosting algorithm and class boosting algorithm as the first layer learner, performing learner training on the training set to obtain the optimal value of the hyperparameters of each learner, and verifying the training result of the learner on the test set; S3: constructing a double-layer learner integrated model: combining two or more of the above-mentioned learners, wherein the hyperparameter values of the learners are the optimal hyperparameter values obtained in step S2; stacking the prediction accuracy of each learner combination on the total data set to form a new data set; using the new data set as the input feature and the measured porosity as the output target, performing model training on the second layer meta-learner to finally construct a double-layer learner integrated model.

2. The dual-layer learner ensemble model for predicting the porosity of a mixed accumulation shale of claim 1, wherein, In step S1, the strength of the correlation between the well logging data and the measured porosity is judged by the following method: drawing a correlation curve of the measured porosity and the well logging data to obtain the correlation coefficient of the curve in the figure, and the greater the correlation coefficient, the stronger the correlation between the well logging data and the measured porosity.

3. The dual-layer learner ensemble model for predicting the porosity of a mixed accumulation shale of claim 1, wherein, In step S1, when performing principal component analysis on the well logging data, the well logging data is normalized to map it to the [0, 1] interval to eliminate the influence of abnormal values in the well logging data on the principal component analysis.

4. The dual-layer learner ensemble model for predicting the porosity of a mixed accumulation shale of claim 1, wherein, In step S2, whether the hyperparameters reach the optimal value is judged by the following method: when the prediction accuracy of the training set reaches the maximum, the corresponding hyperparameter value is the optimal value of the hyperparameters.

5. The dual-layer learner ensemble model for predicting the porosity of a mixed accumulation shale of claim 1, wherein, In step S3, a linear regression function, a K nearest neighbor algorithm, a random forest algorithm, an extreme gradient boosting algorithm, an adaptive boosting algorithm, a light gradient boosting algorithm or a class boosting algorithm is used as a meta-learner.

6. The dual-layer learner ensemble model for predicting the porosity of a mixed accumulation shale of claim 5, wherein, In step S3, a linear regression function is used as a meta-learner.

Images