Marine low-permeability reservoir classification evaluation optimization method based on machine learning

Abstract

The invention provides an offshore low-permeability reservoir classification evaluation optimization method based on machine learning, and relates to the technical field of offshore low-permeability reservoir classification, and the method comprises the following steps: S1, integrating core experiment data, logging curves and production dynamic records of an offshore low-permeability reservoir, extracting a multi-dimensional feature set, removing abnormal data points, and complementing missing parameters; s2, determining an initial feature weight based on grey correlation analysis, introducing a production dynamic feedback mechanism to adjust a parameter contribution degree, and constructing a composite evaluation factor Z containing a sorting coefficient correction term; s3, training subsets are divided according to the permeability, a corresponding regularization constraint intensity training model is adopted, kernel function bandwidth is automatically matched, and a loss function fault-tolerant threshold value is dynamically adjusted; and S4, deploying the model after cross validation, regularly updating a weight coefficient after production, and triggering model retraining through residual monitoring. The method improves the accuracy and reliability of the classification evaluation of the offshore low-permeability reservoir, adapts to the dynamic change of the reservoir, and has a good application prospect.

Description

Technical Field

[0001] This invention relates to the field of marine low-permeability reservoir classification technology, and in particular to a machine learning-based method for optimizing the classification and evaluation of marine low-permeability reservoirs. Background Technology

[0002] For low-permeability offshore reservoirs, traditional classification methods are mainly based on geological characteristics and fluid properties, such as crude oil viscosity, trap type, reservoir properties, and burial depth. Their advantages are simplicity, intuitiveness, and easy parameter acquisition. However, this method relies too heavily on a single parameter, making it difficult to comprehensively evaluate complex heterogeneous reservoirs from multiple factors, and it does not adequately consider dynamic development effects, resulting in poor accuracy.

[0003] To overcome the limitations of single parameters, and to make the method applicable to complex heterogeneous reservoirs, while allowing for dynamic adjustment of classification criteria according to the development stage, a cluster-based classification method is currently used. This method integrates multiple parameters such as porosity, permeability, and oil saturation, and utilizes phylogenetic clustering or the K-Means algorithm to divide the flow units. However, this method has high requirements for data quality and high computational complexity, and the high cost of acquiring offshore data makes this approach difficult to widely apply.

[0004] Therefore, there is an urgent need for a machine learning-based method to optimize the classification and evaluation of low-permeability offshore reservoirs, which addresses the problem of unreliable evaluation indicators caused by high data dimensionality and poor correlation in the classification process of low-permeability offshore oil reservoirs. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based optimization method for classifying and evaluating low-permeability offshore reservoirs. This method aims to avoid the loss of effective information during classification and dimensionality reduction, capture the relationship between feature parameters and evaluation indices through machine learning, and continue learning and verification even when some feature parameters are missing. It is a highly versatile reservoir evaluation method. The various technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a machine learning-based method for classifying, evaluating, and optimizing low-permeability offshore reservoirs, comprising the following steps:

[0008] S1: Data acquisition and preprocessing, integrating core experimental data, well logging curves and production dynamic records of low-permeability offshore reservoirs, extracting multidimensional feature sets, removing abnormal data points and filling in missing parameters;

[0009] S2: Environment construction, based on grey relational analysis to determine the initial feature weights, introduce a production dynamic feedback mechanism to adjust the parameter contribution, and construct a composite evaluation factor Z that includes a sorting coefficient correction term;

[0010] S3: Model building and training. The training subset is divided according to the penetration rate and the model is trained with the corresponding regularization constraint strength. The kernel function bandwidth is automatically matched and the fault tolerance threshold of the loss function is dynamically adjusted.

[0011] S4: Model validation and evaluation. After cross-validation, the model is deployed and the weight coefficients are updated regularly after deployment. Residual monitoring triggers model retraining.

[0012] Preferably, in step S1, an abnormal data point is removed by a geological constraint algorithm based on the porosity-permeability trend line, and missing parameters are filled by Kriging space interpolation.

[0013] Preferably, in step S2, the initial weights are determined by calculating the correlation ratio between each feature and the reservoir quality factor, and the contribution of each parameter is adjusted by the entropy method.

[0014] Preferably, in step S2, the composite evaluation factor Z is the product of the normalized multivariate classification coefficient and the normalized gray classification coefficient.

[0015] Preferably, in step S3, the training subset is divided according to penetration rate into the following categories:

[0016] Ultra-low permeability subset, permeability <0.1mD;

[0017] Low-permeability subset, with a permeability of 0.1-10 mD.

[0018] Preferably, the regularization penalty coefficient range for the ultra-low permeability subset is [0.1, 1.0], and for the low permeability subset it is [1.0, 5.0].

[0019] Preferably, in step S3, the radial basis kernel function bandwidth is determined by an adaptive adjustment method of gamma=scale.

[0020] Preferably, in step S3, the fault tolerance threshold of the loss function is 0.05 when the proportion of missing data is >10%, and 0.01 otherwise.

[0021] Preferably, in step S4, cross-validation is achieved using mean square error, coefficient of determination, and a visual diagnostic chart.

[0022] Preferably, in step S4, when the average absolute error between the predicted oil displacement efficiency and the actual value is >5% or the maximum error is >10%, model retraining is triggered.

[0023] This invention provides a machine learning-based method for classifying and evaluating low-permeability offshore reservoirs. It rigorously selects and preprocesses reservoir geological parameter data to form representative training and independent test sets. The standardized feature data is input into a constructed and optimized support vector regression model. A radial basis function kernel captures the nonlinear relationships between high-dimensional features. During model training, the penalty coefficient and kernel parameters are dynamically adjusted, and cross-validation optimizes the hyperparameter combination. The model output is normalized and truncated to ensure that the predicted values ​​conform to the physical range of the geological parameters. Finally, a visualized evaluation report is generated, including a comparison of actual and predicted values ​​and residual analysis, forming an intelligent prediction system applicable to the evaluation of low-permeability offshore oil reservoirs. Compared with existing technologies, this method retains the information in the parameters to the maximum extent, has higher reliability, adaptively mines potential correlations between data, does not depend on the feature distribution of a specific dataset, and has greater versatility. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the machine learning-based classification, evaluation and optimization method for low-permeability marine reservoirs of the present invention;

[0026] Figure 2 This is a visualization comparison chart of the predicted curve and actual data of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] To address the problems in existing methods for classifying and evaluating low-permeability offshore reservoirs, such as high data dimensionality, poor data processing performance, unreasonable evaluation factor construction, poor correlation, insufficient model training targeting, and unreliable evaluation indicators due to the lack of dynamic update mechanisms, this invention provides a machine learning-based optimization method for classifying and evaluating low-permeability offshore reservoirs. Figure 1 This is a flowchart of this embodiment, as shown below. Figure 1 As shown, the method includes the following steps:

[0029] S1: Data acquisition and preprocessing, integrating core experimental data, well logging curves and production dynamic records of low-permeability offshore reservoirs, extracting multidimensional feature sets, removing abnormal data points and filling in missing parameters;

[0030] S2: Environment construction, based on grey relational analysis to determine the initial feature weights, introduce a production dynamic feedback mechanism to adjust the parameter contribution, and construct a composite evaluation factor Z that includes a sorting coefficient correction term;

[0031] S3: Model building and training. The training subset is divided according to the penetration rate and the model is trained with the corresponding regularization constraint strength. The kernel function bandwidth is automatically matched and the fault tolerance threshold of the loss function is dynamically adjusted.

[0032] S4: Model validation and evaluation. After cross-validation, the model is deployed and the weight coefficients are updated regularly after deployment. Residual monitoring triggers model retraining.

[0033] This machine learning-based method for classifying and evaluating low-permeability offshore reservoirs employs a rigorous screening and preprocessing approach to reservoir geological parameter data, creating representative training and independent test sets. Standardized feature data is input into a pre-constructed and optimized support vector regression model. The radial basis function kernel captures the nonlinear relationships between high-dimensional features, dynamically adjusting the penalty coefficient and kernel parameters during model training. Cross-validation optimizes the hyperparameter combination. The model output is normalized and truncated to ensure predicted values ​​conform to the physical range of geological parameters. Finally, a visualized evaluation report is generated, including a comparison of actual and predicted values ​​and residual analysis. This forms an intelligent prediction system applicable to the evaluation of low-permeability offshore reservoirs. Compared to existing technologies, this method maximizes the retention of information in the parameters, exhibits higher reliability, adaptively mines potential correlations between data, is independent of the feature distribution of specific datasets, and has greater versatility.

[0034] In step S1, a multidimensional feature set is extracted, including a multidimensional feature set reflecting reservoir properties, fluid characteristics, and development potential. Data points that do not conform to reservoir geological patterns are removed using a geological constraint-based anomaly detection algorithm. In this embodiment, a geological constraint algorithm based on reservoir geological parameters such as porosity-permeability trend lines is used to remove abnormal data points that deviate from geological patterns, and Kriging space interpolation is used to complete missing parameters, effectively ensuring data quality and providing a reliable data foundation for subsequent model training.

[0035] In the specific data acquisition and preprocessing process, reliable formation characteristic parameters, such as permeability and porosity, are obtained through techniques such as oilfield geophysical exploration, drilling logging, and core analysis. Comprehensive and reliable data can improve the accuracy and reliability of the model.

[0036] In step S2, the initial weights are determined by calculating the correlation ratio between each feature and the reservoir quality factor, and the contribution of each parameter is adjusted by using the entropy method.

[0037] In this embodiment, after determining the initial feature weights based on grey relational analysis, a dynamic production feedback mechanism is introduced to adjust the contribution of each parameter in real time using the entropy method.

[0038] Specifically, based on the data characteristics of different sections or production stages of the reservoir, the contribution of each parameter can be adjusted in real time using the entropy method to improve the model's predictive accuracy. Considering the heterogeneous characteristics of offshore reservoirs, a sorting coefficient correction term is added to the traditional RQI index to construct a composite evaluation factor Z. The composite evaluation factor Z is the product of the normalized multivariate classification coefficient and the normalized grey classification coefficient.

[0039] By combining grey relational analysis and production dynamic feedback mechanisms, reservoir characteristics can be reflected more comprehensively and accurately, thus improving the scientific nature of the evaluation.

[0040] In the specific data preprocessing process, in order to avoid extreme data affecting the model's data capture, the part with more complete parameters is selected from the original dataset as the basis, and the missing data is filled by interpolation. Some data can be integrated by calculation to reduce the amount of data required for model fitting.

[0041] Formation parameters are categorized into four types based on pore-throat structure, heterogeneity, fluid occurrence, and start-up difficulty. Grey analysis is used to determine the correlation coefficients between evaluation indicators and the Reservoir Quality Index (RQI), objectively and quantitatively reflecting the importance of the selected indicators in reservoir evaluation and identifying four other key controlling factors. The weights of these key controlling factors are determined by the grey relational analysis results, as shown in Formula 1.

[0042]

[0043] In the formula: αi is the weight of each factor, which is dimensionless; ri is the correlation degree of each factor, which is dimensionless.

[0044] Furthermore, based on the positive and negative impacts of various parameters on reservoir space and permeability, the reservoir evaluation factors RQI, main channel throat radius, mobile fluid saturation, and sorting coefficient are positively correlated with effective reservoir space and actual seepage capacity, and negatively correlated with displacement pressure. Using multivariate analysis, the five main controlling factors are normalized and combined, as shown in Formula 2:

[0045]

[0046] In the formula: D is the multivariate classification coefficient, which is dimensionless; RQI is the reservoir quality factor, which is dimensionless;

[0047] RQImax is the maximum reservoir quality factor, dimensionless; So is the mobile fluid saturation, %;

[0048] Somax is the maximum saturation of the movable fluid, %; R is the radius of the main flow throat, μm;

[0049] Rmax is the maximum mainstream throat radius, in μm; F is the sorting coefficient, dimensionless.

[0050] Fmax is the maximum sorting coefficient, dimensionless; P is the exhaust pressure, MPa.

[0051] Pmax is the maximum exhaust pressure, in MPa.

[0052] By combining the normalized multivariate and grey evaluation coefficients, the evaluation index Z describing the comprehensive properties of the reservoir is obtained, as shown in Formula 3:

[0053] Z = D·H (3)

[0054] In the formula: Z is the comprehensive classification coefficient, dimensionless; D is the normalized multivariate classification coefficient, dimensionless; H is the normalized gray classification coefficient, dimensionless.

[0055] In step S3, the samples are divided into two training subsets, ultra-low permeability and low permeability, based on the reservoir permeability distribution characteristics, and different regularization constraint strengths are applied to each subset. During model training, the optimal kernel function bandwidth is automatically matched, and the fault tolerance threshold of the loss function is dynamically adjusted based on data quality.

[0056] Specifically, the training subsets are divided according to permeability into: an ultra-low permeability subset (permeability < 0.1 mD) and a low permeability subset (permeability 0.1-10 mD). The regularization penalty coefficient for the ultra-low permeability subset ranges from [0.1, 1.0], and for the low permeability subset it ranges from [1.0, 5.0]. The radial basis function bandwidth is determined using an adaptive adjustment method of gamma = scale. The loss function tolerance threshold is 0.05 when the missing data ratio is >10%, and 0.01 otherwise.

[0057] This embodiment divides the training subset according to the penetration rate and adopts different regularization constraint strengths. At the same time, it dynamically adjusts the kernel function bandwidth and the fault tolerance threshold of the loss function, making the model training more targeted and improving the model's accuracy and generalization ability.

[0058] In the specific process of building and training the model, the environment required for the support vector machine model is built using Conda, and it is constructed based on the Scikit-learn machine learning framework, the Pandas data processing library, and the NumPy numerical computing tool.

[0059] The hyperparameter optimization penalty coefficient is adjusted according to the reservoir permeability level: when permeability < 0.1 mD, the range is [0.1, 1.0]; when 0.1 mD ≤ permeability < 10 mD, the range is [1.0, 5.0]. The tolerance error is related to data integrity: when the proportion of missing data > 10%, ε = 0.05; otherwise, ε = 0.01. The fault tolerance threshold of the loss function is dynamically adjusted based on data quality.

[0060] In this embodiment, Z is calculated according to Formulas 1, 2, and 3, and then training and testing datasets are created. The feature standardization process uses a standardizer fitted to the training set to uniformly process the test data, preventing data leakage. The preprocessed objects and the model are jointly saved as a pk1 file, forming an end-to-end prediction pipeline.

[0061] In step S4, cross-validation is performed before model deployment to ensure generalization performance, and new data is collected periodically after deployment to update the weight coefficients. A residual monitoring mechanism is established to trigger the model retraining process when the prediction error exceeds a threshold.

[0062] Cross-validation was achieved using mean squared error, coefficient of determination, and visual diagnostic charts.

[0063] Model validation is based on field measured data from the oilfield. When the average absolute error between the predicted oil displacement efficiency and the actual value exceeds 5% or the maximum error exceeds 10%, model retraining is triggered. By establishing a dynamic model update mechanism and promptly triggering retraining through residual monitoring, the model can adapt to dynamic changes in the reservoir, ensuring the long-term validity of the evaluation results.

[0064] In the specific process of model verification and evaluation, field measured data from oilfields are used as the basis. The average absolute error between the predicted oil displacement efficiency and the actual value is ≤5%, the maximum error is ≤10%, and it must be better than the prediction results of the formula data. At the same time, a residual monitoring mechanism is established, and the model retraining process is triggered when the prediction error exceeds the threshold.

[0065] like Figure 2 As shown, in this embodiment, the support vector regression algorithm is called via Python. A radial basis function kernel is set to capture nonlinear relationships. The model complexity is dynamically controlled by a penalty coefficient related to permeability. The kernel bandwidth is adaptively adjusted using gamma = scale. Epsilon is set to be correlated with data integrity to improve robustness to noise and adapt to small sample data. The evaluation module synchronously outputs MSE and R... 2 A visual diagnostic chart is used to perform multi-dimensional validation of the model training. Models with better training results after parameter adjustments are saved for data prediction.

[0066] The table below compares the residuals and variances of the training and test sets in this embodiment.

[0067] training set test set MSE 0.00002 0.00045 R2 0.9994 0.9091

[0068] As shown in the table above, the MSE of both the training and test sets is at a low level, and R... 2 A value close to 1 indicates that the model has excellent fitting performance and good generalization ability.

[0069] To verify the practical application effect of the model, the measured oil displacement efficiency data from the oilfield were compared with the model predictions. The results are shown in the table below:

[0070] Calculated value % Predicted value % Oil displacement efficiency (linear treatment) % 40.9 38.4 35.0 24.6 26.7 32.9 44.6 46.6 37.0 28.4 30.3 36.3 23.3 26.2 35.4 33.3 34.5 32.5 43.3 43.1 34.2 45.7 43.8 36.9 35.3 35.2 34.8 32.5 32.8 36.3 31.2 34.5 33.7 41.0 31.3 33.7 35.3 40.8 32.1 28.7 30.0 34.9 31.1 32.7 28.8 24.4 35.2 32.3

[0071] The comparison results show that the average absolute error between the predicted and actual values ​​is ≤5%, and the maximum error is ≤10%, which is better than the traditional formulaic prediction results and meets the model accuracy requirements.

[0072] This machine learning-based optimization method for classifying and evaluating low-permeability offshore reservoirs improves training efficiency while effectively avoiding overfitting by adjusting model parameters based on experimental data. The method learns a model that captures the potential mathematical correlation between various reservoir characteristic parameters and reservoir evaluation indices, enabling fine classification and evaluation of low-permeability offshore reservoirs and providing more accurate and stable outputs in applications.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based method for classifying, evaluating, and optimizing low-permeability offshore reservoirs, characterized in that, Includes the following steps: S1: Data acquisition and preprocessing, integrating core experimental data, well logging curves and production dynamic records of low-permeability offshore reservoirs, extracting multidimensional feature sets, removing abnormal data points and filling in missing parameters; S2: Environment construction, based on grey relational analysis to determine the initial feature weights, introduce a production dynamic feedback mechanism to adjust the parameter contribution, and construct a composite evaluation factor Z that includes a sorting coefficient correction term; S3: Model building and training. The training subset is divided according to the penetration rate and the model is trained with the corresponding regularization constraint strength. The kernel function bandwidth is automatically matched and the fault tolerance threshold of the loss function is dynamically adjusted. S4: Model validation and evaluation. After cross-validation, the model is deployed and the weight coefficients are updated regularly after deployment. Residual monitoring triggers model retraining.

2. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 1, characterized in that, In step S1, an outlier data point is removed using a geological constraint algorithm based on the porosity-permeability trend line, and missing parameters are filled in using Kriging space interpolation.

3. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 1, characterized in that: In step S2, the initial weights are determined by calculating the correlation ratio between each feature and the reservoir quality factor, and the contribution of each parameter is adjusted by using the entropy method.

4. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 3, characterized in that: In step S2, the composite evaluation factor Z is the product of the normalized multivariate classification coefficient and the normalized gray classification coefficient.

5. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 1, characterized in that: In step S3, the training subset is divided according to penetration rate into the following categories: Ultra-low permeability subset, permeability <0.1mD; Low-permeability subset, with a permeability of 0.1-10 mD.

6. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 5, characterized in that: The regularization penalty coefficient for the ultra-low permeability subset ranges from [0.1, 1.0], while that for the low permeability subset is [1.0, 5.0].

7. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 6, characterized in that: In step S3, the radial basis kernel function bandwidth is determined using the gamma=scale adaptive adjustment method.

8. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 7, characterized in that: In step S3, the fault tolerance threshold of the loss function is 0.05 when the proportion of missing data is >10%, and 0.01 otherwise.

9. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 1, characterized in that: In step S4, cross-validation is achieved using mean squared error, coefficient of determination, and a visual diagnostic chart.

10. The machine learning-based classification, evaluation, and optimization method for low-permeability offshore reservoirs according to claim 9, characterized in that: In step S4, when the average absolute error between the predicted oil displacement efficiency and the actual value is >5% or the maximum error is >10%, model retraining is triggered.

Images