A cross-well logging feature selection method based on distribution difference perception and related device

Abstract

The application discloses a cross-well logging feature selection method and related device based on distribution difference perception, and belongs to the technical field of oil and gas well logging; the method identifies distribution difference significant logging features by constructing inter-well domain classification tasks, and weakens inter-well domain deviation influence based on a generalization performance oriented ablation strategy; the contribution of features to prediction results is quantified by using an explainability analysis method, and group intelligent optimization search is introduced to obtain an optimal feature combination balancing between prediction accuracy, feature stability and quantity. The application can effectively improve the cross-well generalization performance of a pore pressure prediction model, reduce feature redundancy, and is suitable for cross-well pore pressure prediction and related logging interpretation tasks.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas well logging technology, and relates to a method and related device for selecting cross-well logging features based on distribution difference perception. Background Technology

[0002] Pore ​​pressure is a key geomechanical parameter characterizing the formation's mechanical state in oil and gas drilling engineering. Its accurate prediction is crucial for wellbore stability assessment, drilling fluid density window design, wellbore structure optimization, and safety control throughout the drilling process. Traditional pore pressure prediction methods mainly rely on well logging data, such as empirical formulas and equivalent depth methods. These methods are based on specific physical or empirical models and may be effective under specific regional or geological conditions, but their universality and prediction accuracy are often limited, making them difficult to adapt to complex and variable geological environments.

[0003] In recent years, with the rapid development of artificial intelligence and machine learning technologies, machine learning prediction methods have shown great potential in the fields of well logging interpretation and pore pressure prediction. Compared with traditional models, machine learning algorithms, such as random forests and gradient boosting trees, can extract complex nonlinear mapping relationships between features and target variables from massive, high-dimensional historical well logging data, thereby achieving higher fitting accuracy. Researchers typically train prediction models using complete well logging and measured pore pressure data from one or more drilled wells (source wells), hoping to apply them to target wells in adjacent wells or new blocks to predict pore pressure in un-drilled formations and guide drilling operations.

[0004] However, in practical cross-well prediction applications, models trained on historical data often face severe "domain shift" problems, leading to a significant decrease in their generalization performance. Specifically, different wells are influenced by a variety of complex factors, including sedimentary facies distribution, local tectonic stress fields, diagenesis, logging instrument models and calibrations, wellbore environment, and drilling fluid properties. Even when measuring the same type of physical property, the obtained logging curves often exhibit systematic differences in statistical distribution characteristics (such as mean, variance, distribution pattern, and quantile range). This inconsistency in feature distribution causes a significant deviation between the input data distribution encountered by the model in the new well (target well) and the distribution of the source well data on which it was trained. Machine learning models, especially complex models, are prone to overfitting to specific statistical characteristics of the source well data. When faced with target well data of different distributions, the predictions made by the model based on the importance or interaction patterns of the original features will show significant bias, or even fail.

[0005] Existing feature selection techniques, such as filtering based on Pearson correlation coefficients, recursive feature elimination, and feature ranking methods based on built-in model importance, primarily focus on the correlation between features and the target variable or their predictive power within a single data domain. The core optimization goal of these methods is to achieve the best fit on the training set. However, a feature that is highly correlated with the prediction target in the source well and contributes significantly may introduce substantial noise and uncertainty when applied across different wells due to its large distribution differences, becoming a major obstacle to the model's generalization ability. In other words, existing methods lack an explicit perception and evaluation mechanism for the stability of feature distribution across wells, resulting in selected feature subsets that may have strong explanatory power but poor generalization ability, failing to guarantee the model's predictive robustness and reliability in cross-well scenarios.

[0006] Therefore, there is an urgent need in this field for a novel logging feature selection method. This method should not only assess the interpretability of features for the prediction target, but also proactively identify and quantify the distribution differences of features across different wells. Guided by the generalization performance of the final cross-well prediction, it should select a subset of high-value, robust features that have strong predictive power for pore pressure and are relatively stable in distribution across wells. This would alleviate the domain offset problem at its source, construct a more transferable and practical intelligent prediction model for cross-well pore pressure, and improve the scientific rigor and safety of drilling engineering design. Summary of the Invention

[0007] The purpose of this invention is to provide a method and related device for selecting cross-well logging features based on distribution difference perception, so as to solve the technical problem of insufficient generalization ability of cross-well pore pressure prediction in the prior art.

[0008] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for selecting cross-well logging features based on distribution difference perception, comprising the following steps: Obtain well logging feature data of drilled wells, and quantify the distribution differences of each feature among different wells based on a domain classification model to obtain a descending order of DI value sequence; Based on the DI value sequence, features are removed sequentially, and the model generalization performance after each removal is evaluated. The feature subset that achieves the best generalization performance is selected as the initial feature set. For the initial screening feature set, calculate the marginal contribution of each feature to the prediction target; Using the initial screening feature set as the search space, the search process is guided by the marginal contribution of each feature, and the optimal feature subset is searched through a multi-objective optimization algorithm. The optimal feature subset is output, and a prediction model is trained based on the optimal feature subset for cross-well pore pressure prediction.

[0009] Furthermore, the step of acquiring well logging feature data from drilled wells and quantifying the distribution differences of each feature among different wells based on a domain classification model to obtain a descending order sequence of DI values ​​specifically includes: Acquire well logging feature data of drilled wells and divide the well logging feature data of drilled wells into source well data and target well data; The source well samples and the target well samples are labeled as different domains; Construct an inter-well domain classification task and train a domain classification model to distinguish whether a sample belongs to a source well or a target well. Based on the trained domain classification model, the influence of each logging feature on the inter-well domain classification task is quantitatively evaluated and used as the inter-well distinguishing contribution DI. Sort the features in descending order of their DI values ​​to obtain a sequence of DI values.

[0010] Furthermore, the step of removing features sequentially based on the DI value sequence, evaluating the model's generalization performance after each removal, and selecting the feature subset that achieves the optimal generalization performance as the initial feature set specifically includes: Based on the DI value sequence, features are removed one by one, starting from the high-domain differential features; After each feature removal, the pore pressure prediction model is trained using only the remaining feature data from the source well; The generalization performance of the trained pore pressure prediction model was evaluated on the calibration data of the target well. Record the generalization performance index corresponding to different feature subsets, and select the feature subset that makes the generalization performance index reach its maximum value or enter the plateau period as the initial screening feature set.

[0011] Furthermore, the step of calculating the marginal contribution of each feature in the initial screening feature set to the prediction target specifically includes: Based on the initial screening feature set, a pore pressure prediction model is trained on source well data; The marginal contribution of each feature in the initial screening feature set to the prediction result is calculated using the Shapley value-based interpretability analysis method. Calculate the global importance weight of each feature based on its marginal contribution.

[0012] Furthermore, the step of calculating the global importance weight based on the marginal contribution of each feature is specifically calculated using the following formula:

[0013] In the formula, Representation of features The global importance weight; Representation of features In the i Marginal contribution on each sample; N The total number of samples.

[0014] Furthermore, the step of using the initial feature set as the search space, guiding the search process based on the marginal contribution of each feature, and searching for the optimal feature subset through a multi-objective optimization algorithm specifically includes: The feature selection problem is represented as a combinatorial optimization problem of feature subsets, and the feature subsets are encoded using binary vectors; Based on the global importance weight of each feature, the initial population or particles of the optimization algorithm are guided so that features with high prediction contribution have a higher initial selection probability. Define a fitness function that comprehensively considers the prediction performance of the selected feature subset on the target well, the coverage of the feature contribution, and the penalty term for the number of selected features; An optimization algorithm is used to iteratively search for the feature subset that optimizes the fitness function, and this subset is output as the optimal feature subset.

[0015] Furthermore, the expression for the fitness function is:

[0016] In the formula, This represents the generalization performance estimate obtained based on historical calibration samples or known measurement point data from the target well. This represents the coverage of the global importance weights of the selected features, which is the sum of the global importance weights of the selected features; The feature quantity penalty is the ratio of the number of selected features to the total number of features initially screened; α, β, and γ are preset weighting coefficients.

[0017] Secondly, the present invention provides a cross-well logging feature selection system based on distribution difference perception, comprising: The well-to-well domain difference perception module is used to acquire logging feature data of drilled wells and quantify the distribution differences of each feature between different wells based on the domain classification model to obtain a descending order of DI value sequence. The feature ablation screening module is used to remove features sequentially based on the DI value sequence, evaluate the model generalization performance after each removal, and select the feature subset that achieves the best generalization performance as the initial feature set. The feature contribution analysis module is used to calculate the marginal contribution of each feature to the prediction target for the initial screening feature set. The feature combination optimization search module is used to guide the search process based on the marginal contribution of each feature, using the initial screening feature set as the search space, and to search for the optimal feature subset through a multi-objective optimization algorithm. The output module is used to output the optimal feature subset and train a prediction model based on the optimal feature subset for cross-well pore pressure prediction.

[0018] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for selecting cross-well logging features based on distribution difference perception.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for selecting cross-well logging features based on distribution difference perception.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a cross-well logging feature selection method and related apparatus based on distribution difference perception. Feature selection is performed from the perspective of inter-well feature distribution differences. By constructing an inter-well domain classification task, the importance of inter-well logging feature distribution differences is calculated, and features are screened accordingly. This allows for the identification and elimination of features with significant inter-well distribution differences, effectively reducing performance discrepancies caused by distribution inconsistencies in cross-well prediction. A feature ablation mechanism guided by target well generalization performance is used to evaluate model performance on existing calibration intervals, historical logging points, or auxiliary calibration samples of the target well. This optimizes the feature selection process for generalization performance in cross-well application scenarios, avoiding generalization failures caused by screening solely based on source well fitting capabilities. Furthermore, feature contribution analysis results are used to guide the feature combination optimization search, reducing redundant features and improving search efficiency while maintaining cross-well prediction performance, resulting in a better feature subset. In addition, this method is adaptable to models such as XGBoost or other regression models, exhibiting good versatility and scalability, and improving the stability of pore pressure prediction applications in adjacent or new wells. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system of the present invention; Figure 3 This is a heatmap of the feature correlation matrix of an embodiment of the present invention; Figure 4 In the RFECV feature selection process of this embodiment of the invention, R 2 A schematic diagram illustrating the relationship between score and the number of features; Figure 5 This is a schematic diagram of the overall process of the cross-well pore pressure prediction method according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the importance analysis of feature domain differences in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the trend of model prediction performance changes in the feature ablation experiment according to an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the feature marginal contribution analysis based on the SHAP method in an embodiment of the present invention. Figure 9 This is a schematic diagram of the convergence curve of the fitness function during the PSO parameter optimization process in an embodiment of the present invention; Figure 10 This is a comparison chart of prediction results of the RF model across feature selection method in an embodiment of the present invention; Figure 11 This is a comparison chart of prediction results of the XGBoost model across feature selection method in an embodiment of the present invention; Figure 12 This is a comparison chart of prediction results of the LightGBM model across feature selection method in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0026] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0027] The present invention provides a method for selecting cross-well logging features based on distribution difference perception, which can be executed by an electronic device, such as a terminal or a server. The terminal can be a smartphone, tablet, laptop, or other similar device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. It is understood that the present invention does not limit the specific entity executing this method for selecting cross-well logging features based on distribution difference perception.

[0028] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0029] See Figure 1 This invention discloses a method for selecting cross-well logging features based on distribution difference perception, comprising the following steps: S1. Obtain well logging feature data of drilled wells, and quantify the distribution differences of each feature among different wells based on the domain classification model to obtain a descending order of DI value sequence. S101, Obtain well logging feature data of drilled wells, and divide the drilled wells into source wells and target wells, wherein the source wells are modeling wells used to construct pore pressure prediction models, and the target wells are wells to be predicted used to verify the model's cross-well generalization ability; the well logging feature data does not contain prediction target variables.

[0030] S102, mark the source well sample and the target well sample as different domains; mark the source well sample as the first domain (Domain 0) and the target well sample as the second domain (Domain 1).

[0031] S103, Construct an inter-well domain classification task and train a domain classification model (such as a random forest classifier) ​​to distinguish whether a sample belongs to the source well or the target well; S104. Based on the trained domain classification model, the feature discrimination ability evaluation mechanism formed during the training process of the model is used to quantitatively evaluate the influence of each logging feature on the inter-well domain classification task, and the influence is used as the inter-well discrimination contribution (Domain Importance, DI). S105, sort the features in descending order of DI value to obtain the DI value sequence. The larger the DI value, the more significant the distribution difference of the feature among different wells.

[0032] S2, based on the DI value sequence, features are removed sequentially, and the model generalization performance after each removal is evaluated. The feature subset that achieves the best generalization performance is selected as the initial feature set. S201, following the DI value sequence, remove features one by one, starting from the high-domain differential features; S202, after each feature removal, only the remaining feature data of the source well is used to train the pore pressure prediction model (such as XGBoost regression). S203, evaluate the generalization performance of the trained pore pressure prediction model on the calibration data of the target well; specifically, evaluate the generalization performance (such as coefficient of determination R2, mean square error RMSE, etc.) on the existing calibration interval, historical measurement points or auxiliary calibration samples of the target well. S204, record the generalization performance index corresponding to different feature subsets, and select the feature subset that makes the generalization performance index reach the maximum value or enter the plateau period as the initial screening feature set.

[0033] S3, for the initial screening feature set in S2, calculate the marginal contribution of each feature to the prediction target; S301, Based on the initial screening feature set, train the pore pressure prediction model only on source well data; S302, using the interpretability analysis method based on Shapley values, calculate the marginal contribution of each feature in the initial screening feature set to the prediction result; S303, the global importance weight is obtained by averaging the absolute values ​​of the contributions of each feature. SHAP ( f k The specific calculation formula is as follows:

[0034] In the formula, Representation of features The global importance weight; Representation of features In the i Marginal contribution on each sample; N The total number of samples.

[0035] S304 sorts the initial screening features according to SHAP weights, providing guidance for subsequent feature combination optimization.

[0036] S4, using the initial screening feature set as the search space, guide the search process based on the marginal contribution of each feature, and search for the optimal feature subset through a multi-objective optimization algorithm; S401 represents the feature selection problem as a combinatorial optimization problem of feature subsets, and uses a binary vector to encode the feature subsets; 1 indicates that the feature is selected, and 0 indicates that it is not selected.

[0037] S402 guides the initial population or particles of the optimization algorithm based on the global importance weight of each feature, so that features with high prediction contribution have a higher initial selection probability. S403, Define a fitness function that comprehensively considers the predictive performance of the selected feature subset on the target well, the coverage of feature contribution, and a penalty term for the number of selected features; the expression of the fitness function is:

[0038] In the formula, This represents the generalization performance estimate obtained based on historical calibration samples or known measurement point data from the target well. This represents the coverage of the global importance weights of the selected features, which is the sum of the global importance weights of the selected features; The feature quantity penalty is the ratio of the number of selected features to the total number of features initially screened; α, β, and γ are preset weighting coefficients.

[0039] S404. Using optimization algorithms (such as particle swarm optimization, PSO, or genetic algorithm, GA), an iterative search is performed to find the feature subset that makes the fitness function optimal, and the optimal feature subset is output as the output.

[0040] S5, output the optimal feature subset, and train a prediction model based on the optimal feature subset for cross-well pore pressure prediction.

[0041] The final pore pressure prediction model (such as XGBoost or other regression models) is trained using source well data, with the optimal feature subset as input.

[0042] The trained prediction model is applied to the target well data to achieve cross-well pore pressure prediction and output the prediction results.

[0043] See Figure 2 This invention discloses a cross-well logging feature selection system based on distribution difference perception, comprising an inter-well domain difference perception module, a feature ablation screening module, a feature contribution analysis module, a feature combination optimization search module, and an output module. Specifically, the inter-well domain difference perception module acquires logging feature data from drilled wells and quantifies the distribution differences of each feature across different wells based on a domain classification model, obtaining a descending order of DI value sequences. The feature ablation screening module removes features sequentially based on the DI value sequences, evaluates the model generalization performance after each removal, and selects the feature subset that achieves the optimal generalization performance as the initial screening feature set. The feature contribution analysis module calculates the marginal contribution of each feature to the prediction target within the initial screening feature set. The feature combination optimization search module uses the initial screening feature set as the search space, guides the search process based on the marginal contribution of each feature, and searches for the optimal feature subset using a multi-objective optimization algorithm. The output module outputs the optimal feature subset and trains a prediction model based on the optimal feature subset for cross-well pore pressure prediction.

[0044] Example This embodiment uses the prediction of cross-well pore pressure in Block X of an oilfield as an example to illustrate the implementation process of the method of the present invention in detail. Specifically: 1. Data Preparation and Experimental Setup 1.1 Data Source Taking multiple wells within a certain oilfield block as an example, some of these wells are selected as source wells, and the other group is selected as target wells: Source wells: depth range 1100-3200 m, totaling 1800 sets of data, used for model training; Target wells: depth range 1400-3500 m, totaling 2000 sets of data, used for cross-well generalization performance testing.

[0045] 1.2 Well Logging Feature Set There are a total of 20 original well logging features, including: (1) Basic logging parameters: Depth, natural gamma (GR), P-wave transit time (DTc), P-wave velocity (Vp), S-wave velocity (Vs), and bulk density (ρ); (2) Rock mechanical parameters: Poisson's ratio (PR), elastic modulus (E), friction angle (φ), cohesion (C), vertical stress (σv); (3) Derivative characteristics: maximum gamma (GRmax), minimum gamma (GRmin), gamma index (I_GR), and clay content (Vsh); (4) Rock strength parameters: Uniaxial compressive strength of sandstone (UCS Sandstone), uniaxial compressive strength of mudstone (UCSMudstone), and uniaxial compressive strength of shale (UCS Shale). (5) Reference values ​​from literature: Poisson's ratio reference value, elastic modulus reference value.

[0046] The target for prediction is pore pressure Pp (unit: g / cm³). 3 ).

[0047] 1.3 Data Preprocessing The raw logging data is preprocessed as follows: (1) Missing value imputation: Linear interpolation is used to impute missing data to ensure data continuity; (2) Data normalization: The Min-Max normalization method is used to normalize the logging characteristic data to the [0,1] interval. The normalization parameters can be obtained from the source well statistics and applied to the target well to ensure consistent input scale across wells. The formula is:

[0048] 1.4 Comparison Method Settings To verify the effectiveness of the method of the present invention, the following comparison method is set up: (1) Baseline (full feature baseline): Uses all 20 features without feature selection; (2) Pearson correlation analysis: Calculate the Pearson correlation coefficient between each feature and the target variable, and select features with a correlation coefficient greater than 0.6 as input features. See Figure 3 The screening features include: Depth, Vp, Vs, ρ, E, E(Line et al., reference), UCS Sandstone, UCS Mudstone, UCS Shale, σv, C; (3) RFE Recursive Feature Elimination: Recursive feature elimination based on extreme gradient boosting tree importance is adopted, and the optimal number of features is determined by 5-fold cross-validation, such as Figure 4 As shown, four features were ultimately selected: Depth, GR, ρ, and σv. (4) DASP (the method of this invention): a method for selecting cross-well logging features based on distribution difference perception.

[0049] 1.5 Prediction Model The effectiveness of feature selection was validated using three mainstream machine learning regression models: (1) Random Forest (RF); (2) Extreme gradient boosting tree (XGBoost, XGB); (3) Lightweight Gradient Booster (LightGBM, LGB).

[0050] All models were optimized for hyperparameters using grid search combined with 5-fold cross-validation.

[0051] 1.6 Evaluation Indicators The following metrics are used to evaluate predictive performance: (1) Coefficient of determination (R) 2 ): Reflects the goodness of fit between the predicted value and the true value, with a value range of [0,1]. The closer to 1, the better the fit. (2) Mean Square Error (MSE): The average of the squares of the prediction error. The smaller the value, the higher the prediction accuracy. (3) Mean Absolute Error (MAE): The average of the absolute values ​​of the prediction error. The smaller the value, the smaller the prediction deviation. (4) Mean Absolute Percentage Error (MAPE, %): The average of the relative error percentages. The smaller the value, the higher the relative accuracy.

[0052] 2. Step S1: Inter-well domain difference perception modeling like Figure 5 As shown, feature selection is performed according to the overall process of the method of the present invention.

[0053] 2.1 Domain Classification Task Construction The 1800 samples from the source well were labeled as Domain 0, and the 2000 samples from the target well were labeled as Domain 1, constructing an inter-well domain classification task. A total of 3800 samples were merged and used to train the domain classification model.

[0054] 2.2 Domain Classification Model Training A domain classifier was constructed using Randomized Fire (RF). After model training, the domain classifier was able to distinguish the data sources of different wells with high accuracy, indicating that the feature distribution differences between the source well and the target well are extremely significant, further illustrating the necessity of introducing domain difference-aware feature selection in cross-well prediction tasks.

[0055] 2.3 Calculation of the Importance of Domain Differences like Figure 6 As shown, based on the trained domain classification model, the contribution of each feature in distinguishing source wells from target wells is calculated and defined as the Domain Importance (DI) index. The larger the DI value, the more significant the difference in the distribution of that feature between source wells and target wells.

[0056] All 20 features were sorted in descending order according to their DI values. The top 10 features are shown in Table 1, which provides a basis for subsequent feature ablation and selection.

[0057] Table 1 Ranking of the Importance of Feature Domain Differences

[0058] 3. Step S2: Feature ablation screening for generalization performance 3.1 Ablation Strategy In cross-well prediction tasks, features with significant distributional differences can easily introduce domain shift, thereby reducing the model's generalization performance on the target well. Therefore, features with high domain difference importance (DI) are ablated first to mitigate the impact of distributional inconsistencies between wells. Based on the domain difference importance (DI) ranking results obtained in step S1, features are ablated sequentially from high to low DI values. Starting with the feature with the highest domain difference, GRmax, after removing one feature at a time, the XGBoost regression model is trained using only source well data, and the model's generalization performance is evaluated on the target well data.

[0059] 3.2 Ablation process and results like Figure 7 As shown, the change of R² in the target well during the characteristic ablation process.

[0060] 3.3 Determining the Optimal Stopping Point As shown in Table 2, the R² of the target well reaches its peak after the UCS Shale is removed in the ablation step. At this point, 11 features are retained as the initial screening feature set.

[0061] Table 2. Changes in the generalization performance of the target well during characteristic ablation.

[0062] The nine features removed are: GRmax, GRmin, GR, Vsh, I_GR, σv, Depth, C, and UCS Shale.

[0063] The 11 initial screening features retained are: DTc, Vp, Vs, ρ, PR, E, PR1 (Lin et al., reference), E (MPa) (Lin et al., reference), UCS Sandstone, UCS Mudstone, φ.

[0064] 4. Step S3: Feature Contribution Analysis 4.1 Model Training Based on the initial feature set obtained in step S2, a regression prediction model is constructed using only source well data to characterize the mapping relationship between features and prediction targets, providing a foundation for subsequent feature contribution analysis.

[0065] 4.2 Calculation of Feature Contribution The TreeSHAP method, based on Shapley value theory, is used to quantify the marginal contribution of each feature in the model prediction.

[0066] For each feature, the average absolute value of its SHAP values ​​across all source well samples is defined as the global contribution of that feature, and its calculation formula is as follows:

[0067] 4.3 Ranking of Feature Contribution like Figure 8 As shown in Table 3, the SHAP global importance weights of the 11 initially screened features are used to guide the probability initialization in the subsequent optimization search process, so that the search process prioritizes features that contribute highly to the prediction.

[0068] Table 3. Ranking of Feature SHAP Importance

[0069] 5. Step S4: Feature combination optimization search 5.1 Optimization Problem Modeling The feature selection problem is represented as a feature subset combination optimization problem. A binary encoding is used to represent the feature selection scheme, with the encoding vector x = [x1, x2, ..., x11], where xi=1 indicates that the i-th feature is selected, and xi=0 indicates that no feature is selected.

[0070] 5.2 Optimize search methods A particle swarm optimization (PSO) algorithm is used to perform a global search of the feature subset space. Iterative optimization of feature combinations is achieved through particle position updates and fitness evaluation mechanisms.

[0071] 5.3 SHAP-guided probability initialization Based on the feature SHAP weights obtained in step S3, the initial position of the particle is probabilistically guided. The initial probability of the k-th feature being selected is defined as:

[0072] Particles are initialized based on the above probability distribution, so that the particle swarm is more likely to include features with higher prediction contributions in the early stages of the search.

[0073] In this embodiment, the top five features with SHAP weights and their initial selection probabilities are as follows: ρ (32.79%), φ (22.57%), Vp (14.98%), DTc (11.40%) and PR (8.15%).

[0074] After initialization, the number of features of a single particle in the particle swarm is concentrated between 3 and 5.

[0075] 5.4 Fitness Function Design To achieve a balance between feature reduction and prediction performance, a comprehensive fitness function is constructed as follows:

[0076] in: This represents the generalization performance estimate obtained based on historical calibration samples or known measurement point data from the target well. This indicates the coverage of the selected feature SHAP weights, which is the sum of the selected feature SHAP weights; This represents the feature quantity penalty, which is the ratio of the number of selected features to the total number of features initially screened.

[0077] 5.5 Optimize the search process The convergence process of the PSO algorithm is as follows: Figure 9 As shown.

[0078] 5.6 Optimal Feature Combination (1) ρ (bulk density) (2) φ (friction angle) (3) Vp (longitudinal wave velocity) (4) DTc (longitudinal wave time difference) (5) E (elastic modulus) (6) UCS Mudstone (uniaxial compressive strength of mudstone) Compared to the original 20 features, the number of features is reduced to 6, and the feature dimension is reduced by about 70%, which significantly improves the model's generalization ability while maintaining high target well prediction performance.

[0079] 6. Step S5: Cross-well pore pressure prediction and performance comparison 6.1 Feature selection results of the comparison method The feature combinations of the four feature selection methods are shown in Table 4.

[0080] Table 4 Feature selection results of different methods

[0081] 6.2 Performance Comparison of the Three Models Three prediction models (RF, XGB, and LGB) were trained on source well data using selected feature combinations, and their cross-well prediction performance was evaluated on target wells. All models underwent hyperparameter optimization via grid search. Detailed performance results are shown in Table 5.

[0082] Table 5. Cross-well prediction performance of different feature selection methods on three models

[0083] 6.3 Display of Prediction Results The prediction results of the method of this invention on three models are as follows: Figure 10 , Figure 11 and Figure 12 As shown: 6.4 Results Analysis and Discussion (1) Improved cross-well prediction performance: Compared with the Baseline, Pearson and RFE methods, the method of this invention improves the target well prediction accuracy on the three models.

[0084] (2) Advantages in feature dimension and efficiency: The method of this invention reduces the number of features from 20 to 6, and the feature dimension is reduced by 70%, thereby improving computational efficiency.

[0085] (3) Feature stability: The selected feature subset exhibits stable prediction performance across different wells, reducing the impact of well-to-well differences on model performance.

[0086] 7. Technical Effects of this Embodiment This embodiment verifies the implementation process and effects of the method of the present invention based on the cross-well pore pressure prediction task in Block X of an oilfield, and achieves the following technical effects: (1) By analyzing the differences between wells, we can identify and eliminate features with significant differences in distribution between wells, thereby reducing the impact of differences between wells on the prediction results.

[0087] (2) By using feature ablation and optimization search for cross-well generalization performance, the optimal feature combination containing 6 features is obtained, reducing the number of features by 70% compared to the original feature set.

[0088] (3) The target well prediction accuracy is better than the comparison methods on the three models of random forest, XGBoost and LightGBM, which verifies the effectiveness of the method of the present invention in cross-well prediction tasks.

[0089] (4) The selected features are all conventional logging or rock mechanics parameters, which are easy to obtain in engineering and practical application.

[0090] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a cross-well logging feature selection method based on distributed difference sensing.

[0091] This invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the cross-well logging feature selection method based on distribution difference sensing in the above embodiments.

[0092] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for selecting cross-well logging features based on distribution difference perception, characterized in that, Includes the following steps: Obtain well logging feature data of drilled wells, and quantify the distribution differences of each feature among different wells based on a domain classification model to obtain a descending order of DI value sequence; Based on the DI value sequence, features are removed sequentially, and the model generalization performance after each removal is evaluated. The feature subset that achieves the best generalization performance is selected as the initial feature set. For the initial screening feature set, calculate the marginal contribution of each feature to the prediction target; Using the initial screening feature set as the search space, the search process is guided by the marginal contribution of each feature, and the optimal feature subset is searched through a multi-objective optimization algorithm. The optimal feature subset is output, and a prediction model is trained based on the optimal feature subset for cross-well pore pressure prediction.

2. The method for selecting cross-well logging features based on distribution difference perception according to claim 1, characterized in that, The steps of acquiring well logging feature data of drilled wells and quantifying the distribution differences of each feature among different wells based on a domain classification model to obtain a descending order sequence of DI values ​​specifically include: Acquire well logging feature data of drilled wells and divide the well logging feature data of drilled wells into source well data and target well data; The source well samples and the target well samples are labeled as different domains; Construct an inter-well domain classification task and train a domain classification model to distinguish whether a sample belongs to a source well or a target well. Based on the trained domain classification model, the influence of each logging feature on the inter-well domain classification task is quantitatively evaluated and used as the inter-well distinguishing contribution DI. Sort the features in descending order of their DI values ​​to obtain a sequence of DI values.

3. The method for selecting cross-well logging features based on distribution difference perception according to claim 1, characterized in that, The step of removing features sequentially based on the DI value sequence, evaluating the model's generalization performance after each removal, and selecting the feature subset that achieves the optimal generalization performance as the initial feature set specifically includes: Based on the DI value sequence, features are removed one by one, starting from the high-domain differential features; After each feature removal, the pore pressure prediction model is trained using only the remaining feature data from the source well; The generalization performance of the trained pore pressure prediction model was evaluated on the calibration data of the target well. Record the generalization performance index corresponding to different feature subsets, and select the feature subset that makes the generalization performance index reach its maximum value or enter the plateau period as the initial screening feature set.

4. The method for selecting cross-well logging features based on distribution difference perception according to claim 1, characterized in that, The step of calculating the marginal contribution of each feature in the initial feature set to the prediction target specifically includes: Based on the initial screening feature set, a pore pressure prediction model is trained on source well data; The marginal contribution of each feature in the initial screening feature set to the prediction result is calculated using the Shapley value-based interpretability analysis method. Calculate the global importance weight of each feature based on its marginal contribution.

5. The method for selecting cross-well logging features based on distribution difference perception according to claim 4, characterized in that, The step of calculating the global importance weight of each feature based on its marginal contribution is specifically calculated using the following formula: In the formula, Representation of features The global importance weight; Representation of features In the i Marginal contribution on each sample; N The total number of samples.

6. The method for selecting cross-well logging features based on distribution difference perception according to claim 1, characterized in that, The step of using the initial feature set as the search space, guiding the search process based on the marginal contribution of each feature, and searching for the optimal feature subset through a multi-objective optimization algorithm specifically includes: The feature selection problem is represented as a combinatorial optimization problem of feature subsets, and the feature subsets are encoded using binary vectors; Based on the global importance weight of each feature, the initial population or particles of the optimization algorithm are guided so that features with high prediction contribution have a higher initial selection probability. Define a fitness function that comprehensively considers the prediction performance of the selected feature subset on the target well, the coverage of the feature contribution, and the penalty term for the number of selected features; An optimization algorithm is used to iteratively search for the feature subset that optimizes the fitness function, and this subset is output as the optimal feature subset.

7. The method for selecting cross-well logging features based on distribution difference perception according to claim 6, characterized in that, The expression for the fitness function is: In the formula, This represents the generalization performance estimate obtained based on historical calibration samples or known measurement point data from the target well. This represents the coverage of the global importance weights of the selected features, which is the sum of the global importance weights of the selected features; The feature quantity penalty is the ratio of the number of selected features to the total number of features initially screened; α, β, and γ are preset weighting coefficients.

8. A cross-well logging feature selection system based on distribution difference perception, characterized in that, include: The well-to-well domain difference perception module is used to acquire logging feature data of drilled wells and quantify the distribution differences of each feature between different wells based on the domain classification model to obtain a descending order of DI value sequence. The feature ablation screening module is used to remove features sequentially based on the DI value sequence, evaluate the model generalization performance after each removal, and select the feature subset that achieves the best generalization performance as the initial feature set. The feature contribution analysis module is used to calculate the marginal contribution of each feature to the prediction target for the initial screening feature set. The feature combination optimization search module is used to guide the search process based on the marginal contribution of each feature, using the initial screening feature set as the search space, and to search for the optimal feature subset through a multi-objective optimization algorithm. The output module is used to output the optimal feature subset and train a prediction model based on the optimal feature subset for cross-well pore pressure prediction.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for selecting cross-well logging features based on distribution difference perception as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the cross-well logging feature selection method based on distribution difference perception as described in any one of claims 1-7.

Images