A method and system for predicting the CO2 diffusion coefficient of oil-saturated core samples

By employing a two-stage integrated tree regression model and determining the applicable domain for predicting CO2 diffusion coefficients from oil-saturated core samples, the problems of heterogeneity in diffusion data and limited sample size were solved. Stable predictions were achieved under similar media and testing method conditions, improving the accuracy and reliability of the prediction results.

CN123111891APending Publication Date: 2026-11-17UNIV OF SCI & TECH BEIJING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611261906.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for predicting CO2 diffusion coefficients in oil-saturated cores suffer from several problems, including heterogeneous diffusion data sources, a limited number of core target samples, significant bias from direct mixing training, potential target leakage during stacked training, and a lack of reliability indicators for external samples. These issues lead to unstable prediction results.

Method used

By acquiring CO2 diffusion records from multiple sources, and after unifying the units and screening for effectiveness, the data is divided into core core domains and auxiliary domains. Derived features are constructed using continuous technical characteristics, and a two-stage integrated tree regression model is used for training. Combined with the determination of applicable domains, an effective diffusion coefficient prediction is generated.

Benefits of technology

The stability and reliability of CO2 diffusion coefficient prediction from oil-saturated cores have been improved, ensuring the accuracy of the prediction results within the applicable domain. The applicable domain determination restricts the use of the model within the coverage area and outputs reliability prompts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123111891A_ABST
    Figure CN123111891A_ABST
Patent Text Reader

Abstract

This invention relates to the field of oil and gas reservoir development, carbon dioxide flooding, and geological storage parameter prediction technology, and discloses a method and system for predicting the CO2 diffusion coefficient of oil-saturated core samples. The prediction method includes collecting CO2 diffusion records from multiple sources, standardizing units, filtering out invalid and duplicate records, and dividing the core core domain into core and auxiliary domains based on the medium, boundary conditions, testing methods, and model applications. An independent test set is reserved from the core core domain, and the remaining core training samples are randomly or hierarchically divided into K-fold partitions. The first-stage model is trained using auxiliary domain data and K1 subsets of the core core domains. This method utilizes heterogeneous data to learn the overall order of magnitude within the coverage of similar experimental systems and training databases, and controls the leakage of target information in stacked training through out-of-fold prediction. By determining the applicable domain, the output domain is marked as in-domain, boundary domain, or out-of-domain, which can provide diffusion parameters for pressure decay experimental design and CO2 flooding numerical simulation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of oil and gas reservoir development, carbon dioxide flooding and geological storage parameter prediction technology, specifically a method and system for predicting the CO2 diffusion coefficient of oil-saturated core samples. Background Technology

[0002] CO2 diffusion in crude oil and oil-saturated porous media controls crude oil expansion, viscosity reduction, light component extraction, and mass transfer front advancement. It is a key input parameter in numerical simulations of CO2 huff and puff, miscible or near-miscible flooding, and geological sequestration. The effective diffusion coefficient in oil-saturated cores is affected not only by temperature, pressure, and fluid properties, but also by porosity, permeability, core size, boundary conditions, and testing methods. Pressure decay methods, constant pressure methods, nuclear magnetic resonance methods, X-ray CT methods, interface tracking methods, and numerical inversion methods typically require high-temperature, high-pressure equipment and long testing cycles. Different studies use different equations of state, boundary conditions, pressure curve segments, and inversion models, leading to differences in the physical definition, measurement scale, and numerical range of the diffusion coefficient. Existing data-driven methods often directly mix data from different media and testing methods for training, or establish a single-layer mapping using only a limited number of core samples. The former can easily lead the model to misinterpret differences in the experimental domain as core diffusion patterns, while the latter makes it difficult to fully learn the overall order-of-magnitude relationship between temperature, pressure, viscosity, and diffusion coefficient. In two-stage or stacked models, if the fitting results obtained by the model, which includes its own target value, are directly used as inputs for the next stage, target information leakage will occur, resulting in overestimating the validation metrics. Meanwhile, for samples to be predicted that exceed the coverage of training temperature, pressure, viscosity, permeability, or method category, the reliability of the prediction results cannot be guaranteed without determining the applicable domain.

[0003] Therefore, there is a need for a technical solution that does not aim at extrapolation across all experimental systems, but rather limits the prediction stability of CO2 effective diffusion coefficient in oil-saturated cores to similar media, similar testing methods, and similar database coverage conditions. This solution would improve the prediction stability through data role classification, extrapolation-assisted prediction, and core domain correction. Summary of the Invention

[0004] Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention aims to solve the problems of heterogeneous diffusion data sources, limited number of target core samples, large bias in direct mixing training, potential target leakage in stacked training, and lack of reliability indications for out-of-domain samples in similar oil-saturated core experimental systems. It provides effective diffusion coefficient prediction values ​​within the applicable domain for pressure decay experimental design and CO2 flooding numerical simulation.

[0006] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the CO2 diffusion coefficient of oil-saturated core samples, comprising the following steps.

[0007] 1. Obtain CO2 diffusion records from multiple sources, the records including diffusion coefficients and corresponding experimental conditions, fluid properties, core properties, and testing method categories; 2. The records are standardized in terms of units, filtered for validity and deleted for duplicate records. The diffusion coefficient is converted into a logarithmic objective. The records are divided into core core domains and auxiliary domains according to the medium type, boundary conditions, test methods and model applications. 3. Construct derived features based on continuous technical characteristics, and retain the test method category, diffusion type, boundary conditions and experimental domain category to characterize different test and inversion definitions; 4. Reserve an independent test set from the core core domain samples that will not participate in model training and parameter selection. Randomly or hierarchically divide the remaining core core domain training samples into K mutually exclusive subsets. For any subset, use the auxiliary domain training samples and the remaining K subsets... The first-stage model is trained using a subset of the core core area, and the first-stage prediction is generated for this subset. After traversing K subsets, an auxiliary prediction for the training samples of the core core area is formed. 5. Input the technical features of the core core domain training samples and the external auxiliary predictions into the second-stage core domain model for training; use the auxiliary domain training samples and the core core domain training samples to fit the first-stage full model; 6. For the oil-saturated core to be predicted, the auxiliary prediction is generated using the first-stage full model. The auxiliary prediction and the technical characteristics of the oil-saturated core to be predicted are input into the second-stage core domain model to obtain the logarithmic prediction value. The effective diffusion coefficient of the oil-saturated core is then obtained through inverse transformation.

[0008] As a further description of the above technical solution, the experimental conditions and fluid properties include one or more of temperature, pressure, crude oil viscosity and crude oil density; the core properties include one or more of porosity, permeability, core length and core diameter; and the test method categories include one of pressure decay method, constant pressure method, CT method, NMR method, PVT method and numerical inversion method.

[0009] As a further description of the above technical solution, the core core domain includes oil-saturated natural cores, artificial cores, or porous media diffusion records that are the same as or similar to the object to be predicted in terms of medium type, saturated fluid, boundary conditions, or testing methods. The auxiliary domain includes one or more of the following: bulk crude oil diffusion, dynamic diffusion, multiphase diffusion, CO2-containing multi-component gas diffusion, enhanced mass transfer, numerical simulation, and records with effective targets but missing some core properties.

[0010] As a further description of the above technical solution, the derived features include , P / T, L / d, and One or more of them, among which For crude oil dynamic viscosity, Where is the core permeability, P is the experimental pressure, T is the absolute temperature, L is the core length, and d is the core diameter. denoted as core porosity, μ0 as normalized viscosity reference value (1 mPa·s), and k0 as normalized permeability reference value (1 mD).

[0011] As a further description of the above technical solution, the random or hierarchical partitioning maintains the sample ratio of each subset according to at least one of diffusion type, test method category, and experimental domain category, where K is 3 to 10.

[0012] As a further description of the above technical solution, the extrapolation auxiliary prediction of each core core domain training sample is generated by the first-stage model, which does not contain the target value of the core core domain training sample in the training data. The independent test set does not participate in the training of the first-stage model, the second-stage core domain model, or the selection of model parameters.

[0013] As a further description of the above technical solution, the first-stage model and the second-stage core domain model are integrated tree regression models, preferably CatBoost regression models. The model receives oil sample type, diffusion type, experimental domain, boundary conditions and test method category, and handles missing continuous technical features through missing value branching.

[0014] As a further description of the above technical solution, it also includes the determination of the applicable domain: comparing the continuous technical characteristics of the oil-saturated core to be predicted with the feature range, standardized distance, nearest neighbor distance or sample density of the training samples in the core core domain, and checking the number of samples in the core core domain for the test method category of the sample to be predicted.

[0015] As a further description of the above technical solution, when the continuous technical features of the sample to be predicted exceed the preset range, its multivariate distance exceeds the preset threshold, the local sample density is lower than the preset value, or the number of samples of the method category is insufficient, an out-of-domain alarm, a low reliability mark, or an expansion of the prediction interval is output.

[0016] As a further description of the above technical solution, the effective diffusion coefficient is output to the pressure decay experiment design module, the CO2 oil displacement numerical simulation module, or the diffusion process numerical simulation module as a material diffusion parameter.

[0017] As a further description of the above technical solution, the core is evacuated and saturated with crude oil. The core is placed under constant temperature and confining pressure conditions, and CO2 is introduced from the end of the core. The pressure decay curve P(t) is recorded. The experimental diffusion coefficient is obtained through a diffusion inversion model consistent with the experimental boundary conditions. The experimental diffusion coefficient is used to verify the prediction error or the applicable domain threshold.

[0018] As a further description of the above technical solution, it also includes obtaining independent verification values ​​through pressure decay experiments, evacuating the core and saturating it with crude oil, placing the core under constant temperature and confining pressure conditions, introducing carbon dioxide from the end of the core, recording the pressure decay curve P(t), obtaining the experimental diffusion coefficient through a diffusion inversion model consistent with the experimental boundary conditions, and using the experimental diffusion coefficient to verify the prediction error. The experimental diffusion coefficient is output to the pressure decay experiment design module, the carbon dioxide flooding numerical simulation module, or the diffusion process numerical simulation module as a material diffusion parameter.

[0019] The objective of this invention is to perform interpolation and near-field prediction within or similar to the training database coverage, without interpreting random point-level validation results as a necessary generalization performance for any unknown literature, any testing method, or conditions significantly exceeding the training range.

[0020] This invention also provides a CO2 diffusion coefficient prediction system for oil-saturated cores, including a data acquisition module, a cleaning and standardization module, a data role classification module, a derived feature module, a first-stage extrapolation prediction module, a second-stage core domain correction module, an applicable domain determination module, and a diffusion coefficient output module; each module is configured to execute the above-described method.

[0021] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a program that, when executed by the processor, implements the above-described method, and provides a computer-readable storage medium storing the program that, when executed by the processor, implements the above-described prediction method.

[0022] Beneficial effects Compared with existing technologies, this invention provides a method and system for predicting the CO2 diffusion coefficient of oil-saturated core samples, which has the following advantages: 1. The method of the present invention divides the data roles of the core core domain and the auxiliary domain, uses the auxiliary data to learn the overall order of magnitude of diffusion, and avoids treating all experimental systems as completely homogeneous samples.

[0023] 2. Secondly, by generating auxiliary predictions outside the fold through random or hierarchical K-folding, the target values ​​of the same core training samples are allowed to re-enter the second stage without undergoing self-fitting in the first stage, thereby reducing the leakage of target information in stacked training.

[0024] 3. In the second stage, core target domain correction is carried out by combining temperature and pressure, fluid properties, core properties, size and method category, so that the prediction results correspond to the technical conditions of similar oil-saturated cores.

[0025] 4. The applicable domain determines the restriction of the model's use within the database coverage area, and outputs reliability prompts for conditions outside the domain or with low sample density.

[0026] 5. The obtained effective diffusion coefficient can be used as the initial parameter for pressure decay experimental schemes, CO2 oil displacement numerical simulations, and diffusion process models. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall process of the method for predicting the CO2 diffusion coefficient of oil-saturated core samples according to the present invention. Figure 2 This is a diagram of the two-stage training and application architecture of the present invention, which employs random or hierarchical K-fold external auxiliary prediction. Figure 3 This is a schematic diagram of a pressure decay experimental setup used to obtain independent core diffusion data. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for predicting the CO2 diffusion coefficient of oil-saturated core samples, including the following steps: S1. Obtain CO2 diffusion records from multiple sources. Each record must include at least the diffusion coefficient. D And including temperature T ,pressure P Crude oil viscosity μ Crude oil density ρ Porosity φ Penetration rate k Core length L and core diameter d The system utilizes one or more continuous technical features, as well as one or more category technical features from oil sample type, diffusion type, experimental domain, boundary conditions, and test method categories, to complete the construction and role classification of a multi-source database.

[0030] S2. Perform unit standardization, target validity screening, task applicability review, abnormal record verification, and duplicate record deletion on multi-source CO2 diffusion records, and convert the diffusion coefficient into a logarithmic target. : ,in D The units are unified as m ² / s .

[0031] Based on the medium type, boundary conditions, testing methods, and target applications of the diffusion records, the records are divided into a core core domain and an auxiliary domain. The core core domain includes oil-saturated natural cores, artificial cores, or porous media diffusion records that are the same as or similar to the target object in terms of medium type, saturated fluid, testing methods, or boundary conditions. The auxiliary domain includes records that can reflect the overall order of magnitude of CO2 diffusion but are not directly used as samples for the final core performance evaluation, such as bulk crude oil diffusion, dynamic diffusion, multiphase or CO2-containing multi-component gas diffusion, enhanced mass transfer, numerical simulation, and records with effective targets but missing some core properties. Data preprocessing and feature construction are then performed.

[0032] S3. Construct derived features based on continuous technical features. Derived features include... , , P / T , L / d , and φ · One or more of them, among which μ 0 is 1 mPa·s , k 0 is 1 mD Meanwhile, the test method category, diffusion type, boundary conditions, and experimental domain category are retained to characterize different test and inversion definitions.

[0033] S4. Reserve an independent test set from the core core sample. This independent test set will not be used for model training, parameter selection, or extrapolation prediction generation. Divide the remaining core core training samples into K mutually exclusive subsets, either randomly or by stratification based on diffusion type or test method category. For the j-th subset, use the auxiliary domain training samples and the remaining K subsets... The first-stage model is trained on a subset of core core domains, and the first-stage prediction is generated for the j-th subset. After traversing K subsets, the extra-variable auxiliary prediction zOOF of each core core domain training sample is obtained, so that the zOOF of any core training sample is generated by the first-stage model that was not trained using the target value of that sample.

[0034] S5. Combine the original technical features, derived features, category technical features, and extrapolation auxiliary prediction zOOF of the core core domain training samples to train the second-stage core domain model. This allows the second-stage core domain model to learn the target domain bias in similar oil-saturated core systems based on the overall order-of-magnitude estimation of the first stage.

[0035] The first-stage model and the second-stage core domain model can employ an integrated tree regression model capable of handling nonlinear relationships, categorical variables, and missing continuous values; the CatBoost regression model is preferred. The first-stage model is used to learn the overall order of magnitude of the diffusion coefficient, while the second-stage model is used to perform target domain correction by incorporating core properties, size, working conditions, and method categories.

[0036] S6. After completing the extrapolation training, the first-stage full model is refitted using training samples from the auxiliary domain and the core core domain. For the oil-saturated core to be predicted, a feature vector x is constructed according to the same units and feature generation rules as in the training stage. The auxiliary prediction z is generated from the full model in the first stage. Then x With z Input the second-stage core domain model to obtain logarithmic predicted values. and through =10^ Inverse transform yields the effective diffusion coefficient .

[0037] The applicable domain determination step is to compare the continuous features of the core sample to be predicted with the feature range, standardized distance, nearest neighbor distance or sample density of the training database; when the sample to be predicted exceeds the preset range, its distance exceeds the threshold or the number of samples of the method category to which it belongs in the training set is insufficient, an out-of-domain alarm, a low reliability mark or an expansion of the prediction interval is output.

[0038] Independent verification values ​​were obtained through pressure decay experiments. The core was evacuated and saturated with crude oil. The core was placed under constant temperature and confining pressure conditions, and carbon dioxide was introduced from the end of the core. The pressure decay curve P(t) was recorded. The experimental diffusion coefficient was obtained through a diffusion inversion model consistent with the experimental boundary conditions. The experimental diffusion coefficient was used to verify the prediction error. The experimental diffusion coefficient was output to the pressure decay experiment design module, the carbon dioxide flooding numerical simulation module, or the diffusion process numerical simulation module as a material diffusion parameter.

[0039] Example 2: Construction and Role Delineation of Multi-Source Databases 955 CO2 diffusion records were collected from publicly available experimental data and existing experimental records. Units for temperature, pressure, viscosity, density, porosity, permeability, core size, and diffusion coefficient were standardized. Records with empty, non-positive, or misleading diffusion coefficients were deleted, resulting in 938 valid target records. 133 records unsuitable for the current prediction task were excluded, and 19 completely duplicate records were deleted, resulting in 786 modeling records. 285 records of ordinary oil-saturated cores or porous media were designated as the core core domain, while the remaining 501 records of bulk crude oil diffusion, dynamic diffusion, multiphase or CO2-containing multi-component gas diffusion, enhanced mass transfer, and numerical simulation were designated as auxiliary domains. The data roles only determine the training purpose in the model and do not change the original measurements.

[0040]

[0041] The core core domain is selected based on media and boundary conditions that are similar to the object to be predicted. For records with significantly different testing methods, they are not directly considered to be completely equivalent to the target core. Instead, they are expressed conditionally through experimental domain, diffusion type, boundary conditions, and testing method category. When the sample size of a certain method category is too small or it lacks comparability with the target application, it can be moved to an auxiliary domain or excluded.

[0042] Example 3: Data Preprocessing and Feature Construction A unified unit for continuous variables can be set as: temperature. K ,pressure MPa Viscosity mPa·s ,density kg / m³ Porosity decimal, permeability mD Core length cm Core diameter cm diffusion coefficient m² / s Null values ​​in categorical variables are uniformly mapped to the "unknown" category. Missing values ​​of continuous variables are handled by the missing value branch of the model, or imputed only within the training fold using the median. The use of independent test set statistics should be avoided.

[0043] For diffusion coefficient D conduct log 10 Transformations are used to reduce skewness across orders of magnitude. Dimensionless logarithmic forms are applied to cross-order-of-magnitude features such as viscosity and permeability. and and construct P / T , L / d , and φ · Derived features, etc.

[0044] Whether a record is included in the core core domain should be determined based on its medium type, whether it is oil-saturated, boundary conditions, testing methods, diffusion coefficient definition, and comparability with the target application. Records that only have diffusion coefficients but lack key core properties can be retained in the auxiliary domain for the first stage of overall-scale learning and are not directly used for the second stage core domain model performance evaluation.

[0045] Example 4: Training of a randomized or hierarchical K-fold two-stage model First, an independent test set was created from 285 core core domain records, with the remaining samples forming the core training set. The independent test set can be partitioned using a fixed random seed and can be stratified according to diffusion type or test method category to ensure that the proportions of the training and test sets in the major categories are similar. The independent test set is used only for final evaluation and does not participate in missing value statistics, feature selection, model parameter selection, or outlier prediction generation.

[0046] K-fold operations are performed only on the core training set, where K is preferably 5. For the j-th fold, the first-stage model is trained using all auxiliary domain training samples and the remaining 4-fold core training samples, and zOOF is generated for the j-th fold core samples. After traversing 5 folds, each core training sample obtains a first-stage auxiliary prediction generated by the model that was not trained using its target value.

[0047] Both the first-stage model and the second-stage model are based on The regression objective is as follows: The first-stage model receives available temperature, pressure, fluid properties, core properties, size, derived features, and category features; the second-stage model receives the original features, derived features, category features, and zOOF of the core core domain.

[0048] Example 5: Model Parameter Setting and Prediction Performance Verification The first and second stages can use a CatBoost regression model, with the root mean square error as the loss function. Optional parameters range from 300 to 3000 iterations, tree depth from 4 to 10, learning rate from 0.01 to 0.2, and L2 regularization coefficient from 1 to 20. Specific parameters are determined through validation within the training set, with a fixed random seed.

[0049] A reproducible example is set up as follows: K=5, random seed 42; first-stage model iterations 1200, tree depth 6, learning rate 0.03; second-stage model iterations 1000, tree depth 5, learning rate 0.03; categorical variables are directly used as categorical features as input. The example parameters are for illustrating the model implementation; in actual use, they can be adjusted without changing the two-stage and out-of-flight prediction logic.

[0050] After completing the second stage of training, the full model from the first stage is refitted using all auxiliary domain training samples and all core training samples, and then used for the sample z to be predicted. The generation of the model in the second stage still uses the out-of-flight prediction zOOF for training, and does not use the self-fit values ​​obtained from the first stage full model on the core training samples to replace zOOF.

[0051] In one embodiment of 20 sets of randomized point-level repeated validations, the coefficient of determination for the two-stage model was 0.9226 ± 0.0277. log 10 The spatial mean absolute error is 0.1237 ± 0.0171, and the root mean square error is 0.1765 ± 0.0304. These results are intended to illustrate the interpolation and near-field prediction capabilities within the database coverage and similar experimental systems, but are not intended to demonstrate general extrapolation capabilities for arbitrary unknown literature or significantly different experimental methods.

[0052] To evaluate the role of each module, this invention can be compared with a single-stage model that uses only the core domain, a model that directly mixes and trains all records, and a model that uses the fitted values ​​of the training samples themselves as the input for the second stage. Evaluation metrics may include R², MAE, RMSE, relative error distribution, and domain coverage.

[0053] Example 6: Applicable Domain Determination and Prediction Output For the core sample to be predicted, collect or input the temperature, pressure, crude oil viscosity, crude oil density, porosity, permeability, core length, core diameter, and the testing method category, such as pressure decay method, constant pressure method, CT method, or NMR method. The system generates derived features according to the same rules as in the training phase.

[0054] The first stage of the full model generates z based on the features of the sample to be predicted. Second-stage core domain model output Finally obtained =10^ m² / s. The prediction results should include at least the following: The input features used, applicable domain labels, and model version; may further include the prediction interval or the number of nearest neighbor samples.

[0055] The applicable domain determination can be carried out in one or more of the following ways: First, compare each continuous feature of the sample to be predicted with the range of the 1st to 99th percentile of the training set; Second, calculate the standardized Euclidean distance or Mahalanobis distance between the sample to be predicted and the core training samples; Third, calculate the average distance or local sample density to the nearest several training samples; Fourth, check whether the method category and diffusion type to be predicted have reached the preset number of samples in the core training set.

[0056] When the sample to be predicted is in a high-density region of the core training set and the main features are within the preset range, the "in-domain" label is output; when some features are close to the boundary, the "boundary domain" label is output; when multiple key features are out of range, the distance exceeds the threshold, or there are not enough training samples for the category, the "out-of-domain" label is output, and a suggestion is made to increase the pressure decay experiment or adopt a more conservative error range.

[0057] Example 7: Pressure Decay Experiment Verification like Figure 3 As shown, the pressure decay experimental apparatus includes a CO2 gas source, a booster pump, a buffer container, a valve assembly, a constant temperature heating device, a core holder, a pressure sensor, a data acquisition system, a vacuum and saturation unit, and a confining pressure pump.

[0058] The specific process is as follows: After the core is evacuated and fully saturated with crude oil, it is loaded into the core holder and a confining pressure higher than the pore pressure is applied. CO2 is introduced into the end of the core under the set temperature and pressure. The corresponding valve is closed and the pressure decay curve P(t) over time is continuously recorded. The experimental diffusion coefficient is obtained using a diffusion inversion model consistent with the boundary conditions.

[0059]

[0060] The four sets of experiments mentioned above used the same crude oil with a density of 850. kg / m³ The model input viscosity is 20. mPa·s The viscosity input should specify the corresponding temperature; when using reference temperature viscosity or uniformly calibrated viscosity, the viscosity definition should be saved as a data field to ensure that the input meaning is consistent between the training and prediction phases.

[0061] The experimental value and the model prediction value are as follows: Core-1, 9.38 × 10⁻⁶ -9 m² / s With 9.23×10 -9 m² / s Core-7, 1.38×10 -8 m² / s With 1.53×10 -8 m² / s Core-13, 2.87×10 -8 m² / s With 2.92×10 -8 m² / s Core-19, 3.22×10 -8 m² / s With 3.34×10 -8 m² / s. The average relative error of the four sets of results is 4.55. %This experiment is used to verify the consistency between model predictions and independent experimental values ​​under combined operating conditions, and is not interpreted as an independent effect of a single factor such as temperature, pressure, porosity, or permeability.

[0062] Example 8: Systems and Electronic Equipment It includes a data acquisition module, a cleaning and standardization module, a data role classification module, a derived feature module, a first-stage out-of-range prediction module, a second-stage core domain correction module, an applicable domain determination module, and a diffusion coefficient output module; each module is configured to execute the prediction methods of Examples 1 to 6 above.

[0063] The prediction system can be deployed on workstations, servers, or experimental data acquisition computers. The data acquisition module receives experimental parameters from database records or data acquisition system outputs. The cleaning and standardization module performs unit unification, validity screening, and target transformation. The data role classification module determines the core core domain and auxiliary domain according to the medium, method, and boundary conditions.

[0064] The first-stage out-of-fold prediction module performs random or hierarchical K-folding within the core training set and generates zOOF; the second-stage core domain correction module combines core core features and zOOF to complete training; the applicable domain determination module outputs domain-in-domain, boundary-domain, or out-of-domain labels based on range, distance, or sample density; the diffusion coefficient output module will... Material parameter interfaces for writing experimental design files, databases, or numerical simulation software.

[0065] It also includes an electronic device consisting of a processor and a memory, wherein the memory stores a program that, when executed by the processor, implements the above-described method; and provides a computer-readable storage medium storing a program.

[0066] The memory can store unit conversion rules, target transformation rules, derived feature rules, category dictionary, training set statistics, model parameters, applicable domain thresholds, and model version. When the processor executes the program in the memory, it implements the above prediction method steps.

[0067] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the CO2 diffusion coefficient of oil-saturated core samples, characterized in that, Includes the following steps: Step 1: Obtain CO2 diffusion records from multiple sources, including diffusion coefficients. D The database is constructed and roles are assigned based on a variety of continuous technical characteristics consisting of experimental conditions, fluid properties, and core properties, as well as various categories of technical characteristics including oil sample type, diffusion type, experimental domain, boundary conditions, and test method categories. Step 2: Perform unit unification, target validity screening, task applicability review, abnormal record verification, and duplicate record deletion on the diffusion records. Convert the diffusion coefficient into a logarithmic objective y. Based on the medium type, boundary conditions, test methods, and target uses corresponding to the diffusion records, divide the records into core core domains and auxiliary domains. Step 3: Construct derived features based on continuous technical features, and retain the test method category, diffusion type, boundary conditions and experimental domain category to characterize different test and inversion definitions; Step 4: Reserve an independent test set from the core core domain samples, and use the remaining core core domain samples as core core domain training samples. Randomly or hierarchically divide these samples into K mutually exclusive subsets. For any subset, use the auxiliary domain training samples and the remaining K subsets... The first-stage model is trained using a subset of core core domains, and a first-stage core domain prediction model is generated for this subset. After traversing K subsets, an auxiliary prediction for the training samples of the core core domains is formed. Step 5: Input the technical features of the core core domain training samples and the external auxiliary predictions into the second-stage core domain model for training, and use the auxiliary domain training samples and the core core domain training samples to fit the first-stage full model. Step 6: For the oil-saturated core to be predicted, use the first-stage full model to generate an external auxiliary prediction. Input the external auxiliary prediction and the technical characteristics of the oil-saturated core to be predicted into the second-stage core domain prediction model to obtain the logarithmic prediction value. Then, through inverse transformation, obtain the effective diffusion coefficient of the oil-saturated core.

2. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: The experimental conditions and fluid properties include one or more of temperature, pressure, crude oil viscosity, and crude oil density; the core properties include one or more of porosity, permeability, core length, and core diameter; and the testing methods include one of pressure decay method, constant pressure method, CT method, NMR method, PVT method, and numerical inversion method.

3. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: The core core domain includes oil-saturated natural cores, artificial cores, or porous media diffusion records that are the same as or similar to the target object in terms of medium type, saturated fluid, boundary conditions, or testing methods. The auxiliary domain includes one or more of the following: bulk crude oil diffusion, dynamic diffusion, multiphase or CO2-containing multi-component gas diffusion, enhanced mass transfer, numerical simulation, and records with effective targets but missing some core properties. Data preprocessing and feature construction are completed, and the logarithmic target... ,in D The units are unified as m² / s .

4. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: The derived features include , P / T, L / d, and One or more of them, among which For crude oil dynamic viscosity, Where is the core permeability, P is the experimental pressure, T is the absolute temperature, L is the core length, and d is the core diameter. denoted as core porosity, μ0 as normalized viscosity reference value (1 mPa·s), and k0 as normalized permeability reference value (1 mD).

5. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: The random or stratified partitioning maintains the sample proportion of each subset according to at least one of diffusion type, test method category, and experimental domain category, with K ranging from 3 to 10. The auxiliary prediction of the training sample of each core core domain is generated by the first-stage core domain prediction model, which does not include the target value of the training sample of that core core domain in the training data. The independent test set does not participate in the training of the first-stage core domain prediction model and the second-stage core domain prediction model, nor in the selection of model parameters. Both the first-stage core domain prediction model and the second-stage core domain prediction model are integrated tree regression models, specifically CatBoost regression models. The integrated tree regression model receives oil sample type, diffusion type, experimental domain, boundary conditions, and test method category, and handles missing continuous technical features through missing value branches.

6. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: It also includes the applicable domain determination step: S1. Compare the continuous technical characteristics of the oil-saturated core to be predicted with the characteristic range, standardized distance, nearest neighbor distance or sample density of the training samples of the core core domain; S2. Check the number of samples in the core core domain for the test method category of the sample to be predicted; S3. When the continuous technical features of the sample to be predicted exceed the preset range, its multivariate distance exceeds the preset threshold, the local sample density is lower than the preset value, or the sample of the method category is insufficient, output an out-of-domain alarm, a low reliability mark, or expand the prediction interval.

7. The method for predicting the CO2 diffusion coefficient of oil-saturated cores according to claim 1, characterized in that: It also includes obtaining independent verification values ​​through pressure decay experiments, evacuating the core and saturating it with crude oil, placing the core under constant temperature and confining pressure conditions, introducing carbon dioxide from the end of the core, recording the pressure decay curve P(t), obtaining the experimental diffusion coefficient through a diffusion inversion model consistent with the experimental boundary conditions, and using the experimental diffusion coefficient to verify the prediction error.

8. A system for predicting the CO2 diffusion coefficient of oil-saturated core samples, characterized in that, It includes a data acquisition module, a cleaning and standardization module, a data role classification module, a derived feature module, a first-stage out-of-range prediction module, a second-stage core domain correction module, an applicable domain determination module, and a diffusion coefficient output module; each module is configured to perform the prediction method according to any one of claims 1 to 7.

9. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program, which, when executed by the processor, causes the processor to perform the prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the prediction method according to any one of claims 1 to 7.