A method and system for predicting crude oil miscibility pressure using gradient-constrained physical information networks

By constructing an automatically differentiable minimum miscibility pressure prediction network, and combining data loss, empirical interval loss, and gradient prior loss, and training in stages, the problem of prediction results deviating from empirical range and violating physical laws in existing technologies is solved, and higher accuracy and stability of crude oil miscibility pressure prediction are achieved.

CN122865262BActive Publication Date: 2026-11-10CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611376918.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-07
Publication Date
2026-11-10
Estimated Expiration
2046-09-07

AI Technical Summary

Technical Problem

When predicting the minimum miscibility pressure of crude oil, existing technologies often produce predictions outside the empirical range due to data-driven models, with local response directions violating physical understanding, and fixed constraint weights affecting the model's insufficient convergence.

Method used

A gradient-constrained physical information network method is adopted. By acquiring the analysis of target crude oil and gas components, an automatically differentiable minimum miscibility pressure prediction network is constructed. A joint objective function is constructed using data loss, empirical interval loss and gradient prior loss. The network is trained in stages. By combining empirical models and physical constraints, the prediction results are ensured to be within a reasonable range.

Benefits of technology

It improves the prediction accuracy, stability and physical consistency under complex injection gas and finite sample conditions, reduces the inhibition of data learning by strong physical constraints in the early stage of training, and achieves more accurate minimum mixing pressure prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122865262B_ABST
    Figure CN122865262B_ABST
Patent Text Reader

Abstract

This invention relates to the field of crude oil extraction technology, specifically to a method and system for predicting crude oil miscibility pressure using a gradient-constrained physical information network. The method involves acquiring the composition of the target crude oil, the composition of the injected gas, and the reservoir temperature, performing feature processing, and then inputting these data into a pre-trained prediction model. The model is constructed using historical miscibility experimental samples, with the measured minimum miscibility pressure as the supervision label, and setting a physical safety interval based on an empirical model. Ordinary partial derivatives are calculated for non-component features, and closed directional derivatives are calculated for component features that satisfy the mole fraction closure condition. This forms a gradient prior loss, which is then dynamically weighted and jointly trained with the data loss and the empirical interval loss in stages. This invention can obtain prediction results without requiring the measured minimum miscibility pressure of the target crude oil and improves the physical consistency and stability of predictions under complex injected gas conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of crude oil extraction technology, and in particular to a method and system for predicting crude oil miscibility pressure using a gradient-constrained physical information network. Background Technology

[0002] In oilfield enhanced oil recovery technologies, CO2 miscible flooding can reduce crude oil viscosity, improve mobility, and increase displacement efficiency. Minimum miscibility pressure (MMP) is a crucial parameter for determining whether injected gas and crude oil can achieve dynamic miscibility, identifying the injection pressure window, and designing capillary experiments. Accurate prediction of the minimum miscibility pressure is of great significance for selecting injection schemes and optimizing reservoir development plans.

[0003] Currently, the minimum miscibility pressure can be determined through capillary experiments, interfacial tension experiments, equation of state calculations, or empirical correlations. Experimental methods are time-consuming and costly; equation of state calculations depend on complete components and parameter adjustments; empirical correlations are convenient, but their applicability is limited by the fitted samples and formula form, making it difficult to fully describe complex injection gases containing N2, H2S, CH4, and intermediate hydrocarbon components.

[0004] Conventional machine learning models can learn the nonlinear relationship between minimum miscibility pressure and crude oil composition, injected gas composition, and reservoir temperature. However, model training typically focuses solely on prediction error. When the sample size is limited, the sample distribution is uneven, or the target operating condition is located at the sample boundary, the prediction results may deviate from the empirically reasonable range, or the local response direction of the predicted pressure to temperature, molecular weight of heavy components, and injected gas composition may be inconsistent with known physical laws.

[0005] If empirical formulas are used directly as training labels or hard constraints, the systematic bias of the empirical formulas will be passed on to the data-driven model. If only empirical values ​​are used as additional inputs, the reasonable range of the network output cannot be guaranteed. If only the range of predicted values ​​is restricted, the direction of change of the network in the local neighborhood cannot be restricted.

[0006] Therefore, there is a need for a method for predicting the minimum miscibility pressure of crude oil that can simultaneously utilize empirical range knowledge and monotonic response knowledge, and can reconcile data fitting with the strength of physical constraints. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting crude oil miscibility pressure by integrating gradient-constrained physical information networks. This aims to solve the technical problems in existing technologies, such as data-driven models easily generating predictions outside the empirical range, local response directions violating physical understanding, and fixed constraint weights affecting model convergence.

[0008] To achieve the above objectives, this invention employs a crude oil miscibility pressure prediction method that integrates gradient-constrained physical information networks, comprising the following steps: The target crude oil with unknown minimum miscibility pressure, the target gas used to inject the target crude oil for oil displacement, and the target reservoir temperature are obtained respectively. The target crude oil and target gas are subjected to component analysis and characterization to determine the target necessary parameters for predicting the minimum miscibility pressure. The target minimum miscibility pressure for the target crude oil and target gas to be miscible is determined by a pre-built target minimum miscibility pressure prediction model based on the target necessary parameters. The physical safety range is determined based on the necessary parameters of the sample and the pre-fitted empirical minimum miscibility pressure model. An automatically differentiable minimum miscibility pressure prediction network is constructed, and derivatives are calculated for non-component monotonic features and component monotonic features expressed in mole fraction. A joint objective function is constructed based on data loss, empirical interval loss, and gradient prior loss, and trained in at least two training phases with physical constraint weights increasing from weak to strong.

[0009] The construction process of the target minimum miscibility pressure prediction model includes: obtaining a training sample set formed by historical miscibility experiments, each training sample including sample crude oil, sample gas, sample reservoir temperature and measured minimum miscibility pressure as a supervision label; The construction process of the target minimum miscibility pressure prediction model includes data prediction processing, empirical interval constraint processing, gradient constraint processing, and joint training processing.

[0010] The data prediction processing is used to input the standardized sample necessary parameters into an automatically differentiable minimum miscibility pressure prediction network, and the minimum miscibility pressure prediction network outputs the sample predicted minimum miscibility pressure. The empirical interval constraint processing is used to calculate the sample empirical minimum miscibility pressure and physical safety interval based on the pre-fitted empirical minimum miscibility pressure model, and to form an empirical interval loss based on the extent to which the sample predicted minimum miscibility pressure exceeds the physical safety interval. The gradient constraint processing is used to calculate the ordinary partial derivative or the closed-formation direction derivative corresponding to the minimum mixing pressure predicted by the sample, and to form a gradient prior loss based on the difference between the corresponding derivative and the preset physical monotonic direction. The joint training process is used to update the network parameters of the minimum mixing pressure prediction network based on data loss, empirical interval loss, and gradient prior loss.

[0011] After training, the minimum mixing pressure prediction network with determined network parameters is saved as the target minimum mixing pressure prediction model. The empirical minimum mixing pressure model, physical safety interval, and gradient constraints are used to train the minimum mixing pressure prediction network or evaluate the physical consistency of the prediction results, and are not used as independent prediction sub-models for outputting the target minimum mixing pressure.

[0012] The expression for the empirical minimum miscibility pressure model is as follows: (1) in, The empirical minimum miscibility pressure is used for the sample and is not included as the final output of the target minimum miscibility pressure prediction model. Tr represents the reservoir temperature, and MWC is used for the measurement. 7+ Crude oil C 7+ Average molecular weight, X vol X is the sum of the molar percentages of N2 and CH4 in crude oil. int The sum of the molar percentages of CO2, H2S, and C2~C6 in crude oil, Tc,CO2 is the critical temperature of CO2, and y i T represents the mole fraction of the i-th component in the injected gas. ci Let be the critical temperature of the i-th component, and α, β, λ, δ, and γ be fitting parameters determined by multiple nonlinear regression based on historical miscibility experimental samples, with the measured minimum miscibility pressure as the dependent variable; the upper and lower bounds of the physical safety interval are respectively... and , The allowable deviation coefficient ε is determined based on the coverage rate of the measured minimum miscibility pressure in the independent validation samples falling within the physical safety range, which reaches a preset coverage threshold. The derivatives of the non-component monotonic characteristics and the component monotonic characteristics expressed as mole fractions are calculated using the automatically differentiable minimum miscibility pressure prediction network described above: For non-component type monotonic features, calculate the partial derivative of the sample predicted minimum miscibility pressure with respect to the monotonic features; For component-type monotonic characteristics expressed in terms of mole fraction, a closed compositional perturbation is constructed while keeping the sum of the mole fractions of each component in the same compositional system constant, and the directional derivative of the sample predicted minimum miscibility pressure along the direction of the closed compositional perturbation is calculated.

[0013] Among them, monotonic features include non-component monotonic features and component monotonic features; for reservoir temperature Tr and crude oil C 7+ Average molecular weight (MWC) 7+ Calculate the partial derivative of the predicted minimum miscibility pressure with respect to the corresponding characteristic; for the monotonic characteristic of the injected gas component type, while keeping the sum of the mole fractions of each component of the injected gas at 1, increase the mole fraction of the target component and correspondingly decrease the mole fraction of at least one other injected gas component, thereby constructing a closed composition perturbation direction and calculating the derivative of the closed composition direction; the injected gas N2 content N 2g and CH4 content C 1g Corresponding to the positive closed-circuit composition direction, the injected gas CO2 content CO 2gThe negative closed composition direction corresponds to the H2S content H2Sg; the ordinary partial derivatives of the non-component type monotonic characteristics and the closed composition direction derivatives of the component type monotonic characteristics are uniformly defined as the unified gradient response quantity; for positive monotonic relations, the square of the negative part of the unified gradient response quantity is used to determine the individual gradient penalty, and for negative monotonic relations, the square of the positive part of the unified gradient response quantity is used to determine the individual gradient penalty.

[0014] The construction process of the empirical minimum miscibility pressure model includes: obtaining the necessary parameters of the sample; analyzing the influence of reservoir temperature, molecular weight of heavy components in crude oil, ratio of volatile components to intermediate components in crude oil, and pseudocritical temperature of mixed injection gas on the measured minimum miscibility pressure based on the necessary parameters of the sample; constructing an initial empirical minimum miscibility pressure model based on the influence relationship; fitting the parameters of the initial empirical minimum miscibility pressure model based on the training sample, and determining the allowable deviation coefficient based on the coverage of validation samples that are independent of the training sample; and excluding the test sample from the parameter fitting, allowable deviation coefficient determination, and model selection process of the empirical minimum miscibility pressure model.

[0015] Among them, the data loss is determined based on the minimum miscibility pressure predicted from the samples and the supervision labels; The empirical range loss is determined based on the extent to which the predicted minimum miscibility pressure exceeds the physical safety range. The gradient prior loss is determined based on the difference between one of the derivatives, either the partial derivative or the directional derivative of the closure-formed perturbation direction, and the preset physical monotonic direction. A joint objective function is constructed based on data loss, empirical interval loss, and gradient prior loss. The joint objective function is: ; Training the minimum mixing pressure prediction network includes a data-dominated phase, a physical constraint asymptotic enhancement phase, and a physical constraint stabilization phase, maintaining stability during the data-dominated phase. Keeping the first empirical interval loss weight and the first gradient prior loss weight unchanged, they are continuously increased during the asymptotic enhancement phase of physical constraints. and During the stable phase of physical constraints, the increase remains. and constant.

[0016] The necessary parameters for the sample include the mole fractions of CO2, N2, H2S, CH4, and C2-C6 in the injected gas; the polar dissolved gas complex characteristics (DAG) of the crude oil; the volatile component complex characteristics (Vol) of the crude oil; and the C2-C4 content, C5-C6 content, and C6 content of the crude oil. 7+ Average molecular weight (MWC) 7+And reservoir temperature Tr; where Vol is the sum of the molar percentages of CH4 and N2 in crude oil; the minimum miscibility pressure prediction network is a residual fully connected neural network using continuously differentiable activation functions.

[0017] This invention also provides a crude oil miscibility pressure prediction system integrating gradient-constrained physical information networks, comprising a parameter acquisition module, a parameter analysis module, a physical interval construction module, a prediction model construction module, a gradient constraint module, a joint training module, and a miscibility pressure prediction module; wherein: The parameter acquisition module is used to acquire a training sample set formed by historical miscibility experiments during the model building stage, and to acquire the target crude oil, target gas and target reservoir temperature with unknown minimum miscibility pressure during the target prediction stage. The training sample set includes the measured minimum miscibility pressure as a supervision label, and the measured minimum miscibility pressure of the target crude oil does not need to be acquired during the target prediction stage. The parameter analysis module is used to perform component analysis and feature processing on sample crude oil and sample gas during the model building stage, determine the necessary parameters of the sample and the corresponding standardized parameters, and determine the necessary parameters of the target in the target prediction stage according to the same feature definition as in the model building stage.

[0018] The physical range construction module is used to determine the empirical minimum miscibility pressure and the physical safety range based on the empirical minimum miscibility pressure model. The prediction model construction module is used to construct a minimum miscibility pressure prediction network that takes standardized sample necessary parameters as input and the sample predicted minimum miscibility pressure as scalar output. The gradient constraint module is used to calculate the ordinary partial derivative of the predicted minimum miscibility pressure for non-component monotonic features, calculate the closed composition direction derivative of the component monotonic features to keep the sum of the mole fractions of each component in the corresponding composition system equal to 1, and determine the gradient prior loss based on the difference between one of the ordinary partial derivative and the closed composition direction derivative and the preset physical monotonic direction. The joint training module is used to construct a joint objective function based on data loss, empirical interval loss and gradient prior loss, and update the network parameters of the minimum mixing pressure prediction network using the joint objective function in at least two training stages according to the physical constraint weights from weak to strong. The minimum mixing pressure prediction network with determined network parameters and its corresponding standardized parameters are then saved as the target minimum mixing pressure prediction model. The miscibility pressure prediction module is used to determine the target necessary parameters according to the feature definition adopted in the model construction stage, to standardize the target necessary parameters using the standardized parameters corresponding to the target minimum miscibility pressure prediction model, to input the standardized target necessary parameters into the target minimum miscibility pressure prediction model for forward calculation, and to determine the predicted minimum miscibility pressure output by the target minimum miscibility pressure prediction model as the target minimum miscibility pressure.

[0019] The present invention also provides a crude oil miscibility pressure prediction device that integrates gradient-constrained physical information networks, including a processor, a user interface, and a memory. The processor, the user interface, and the memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the crude oil miscibility pressure prediction method that integrates gradient-constrained physical information networks.

[0020] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the crude oil miscibility pressure prediction method using a fused gradient constraint physical information network.

[0021] This invention discloses a method and system for predicting crude oil miscibility pressure using a gradient-constrained physical information network. By converting an empirical model into a physically safe range that only applies when limits are exceeded, and by transforming the monotonic relationship between temperature, crude oil properties, and injected gas composition into an automatically differentiable gradient prior, the system can constrain the macroscopically reasonable range and local response direction of the predicted values. Furthermore, by dynamically adjusting the loss weights in stages, the system can reduce the inhibitory effect of strong physical constraints on data learning during the initial training phase. This improves the prediction accuracy, stability, and physical consistency under complex injected gas conditions and limited sample conditions. Attached Figure Description

[0022] 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.

[0023] Figure 1 This is a schematic flowchart of the crude oil minimum miscibility pressure prediction method of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the construction of the target minimum miscibility pressure prediction model of the present invention.

[0025] Figure 3This is a schematic diagram illustrating the relationship between the empirical minimum miscibility pressure and the physical safety range provided by an embodiment of the present invention; Figure 4 This is a schematic diagram of monotonic gradient constraints combining ordinary partial derivatives and closed-form directional derivatives according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the phased dynamic loss weight provided in an embodiment of the present invention; Figure 6 This is a comparison chart of the prediction results of a homogeneous deep neural network and a physical information neural network in the same verification trade-off, according to an embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram of the structural principle of the crude oil miscibility pressure prediction system based on the fusion gradient-constrained physical information network of the present invention.

[0027] Figure 8 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0029] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0031] Figure 1A flowchart of a crude oil minimum miscibility pressure prediction method provided by an embodiment of the present invention is shown. This embodiment can be applied to the prediction of minimum miscibility pressures of crude oil in the context of pure CO2, CO2 mixtures containing N2 or H2S, and hydrocarbon-rich injection gas. The method can be executed by a crude oil minimum miscibility pressure prediction device, which can be implemented using software and / or hardware. Figure 1 The online target prediction stage, as shown, predicts the minimum miscibility pressure of a target crude oil when the minimum miscibility pressure is unknown, without requiring the input of the measured minimum miscibility pressure of the target crude oil. This method includes: S101, obtain the target crude oil with unknown minimum miscibility pressure, the target gas used to inject the target crude oil for oil displacement, and the target reservoir temperature.

[0032] S102, perform component analysis and feature processing on the target crude oil and target gas, determine the necessary parameters of the target according to the feature definition adopted in the model building stage, and process the necessary parameters of the target using the standardized parameters saved in the model building stage in the same standardization method as the training samples.

[0033] S103, the standardized target necessary parameters are input into the pre-constructed target minimum mixing pressure prediction model for forward computation. The target minimum mixing pressure prediction model is a minimum mixing pressure prediction network trained with a joint objective function and whose network parameters are already determined. Let the standardized target necessary parameters be... The network parameters determined after training are: The output of the target minimum miscibility pressure prediction model is then expressed as: (2) In the formula, This represents the minimum miscibility pressure prediction network after the network parameters are determined. This represents the predicted minimum miscibility pressure corresponding to the target crude oil and the target gas.

[0034] S104, output the target minimum miscibility pressure prediction model The target minimum miscibility pressure is determined. The measured minimum miscibility pressure of the target crude oil does not need to be input during the target prediction stage. Optionally, an empirical minimum miscibility pressure and a physical safety range are calculated based on the necessary parameters of the target, and the ordinary partial derivatives or closed-form directional derivatives corresponding to each preset monotonic feature are calculated to evaluate the physical consistency of the prediction results; the empirical minimum miscibility pressure and physical safety range do not replace the output of the target minimum miscibility pressure prediction model.

[0035] Figure 2 The offline construction process of the target minimum miscibility pressure prediction model is shown. The offline stage uses historical experiment labels to train the model, and the construction process includes: S201, Obtain a training sample set formed by historical miscibility experiments. Each training sample includes sample crude oil, sample gas used to inject sample crude oil for oil displacement, sample reservoir temperature, and measured minimum miscibility pressure as a supervision label.

[0036] The sample crude oil can be a historical crude oil sample that has undergone capillary experiments, effervescence experiments, or other miscibility experiments. The sample gas can be pure CO2 or a mixture containing N2, H2S, CH4, or C2-C6 components. The measured minimum miscibility pressure is only used as a supervisory label for offline training samples, preferably derived from historical records with clearly defined experimental temperatures, oil and gas compositions, and judgment criteria, and is not an input requirement for the target crude oil to be predicted.

[0037] The system performs unit standardization, duplicate record identification, component normalization checks, key field integrity checks, and physical range checks on raw data. Records lacking key inputs or measured minimum miscibility pressures are discarded; for convertible temperatures, pressures, and mole fractions, they are uniformly converted to preset units.

[0038] S202 involves performing component analysis and characterization on the sample crude oil and sample gas to determine the necessary sample parameters for predicting the minimum miscibility pressure.

[0039] The essential parameters for the sample include at least the reservoir temperature, crude oil composition parameters, and gas composition parameters. The crude oil composition parameters and gas composition parameters should each include at least the mole fraction of volatile components, the mole fraction of light or intermediate components, and the characterization parameters of heavy components.

[0040] In one specific embodiment, 664 valid records were obtained from 971 original records. Candidate parameters were analyzed using correlation, mutual information, recursive feature elimination, embedded importance, and reservoir engineering knowledge to determine 11 essential parameters for the samples, including CO2g, N2g, H2Sg, C1g, C2g~C6g, DAG, Vol, C2o~C4o, C5o~C6o, and MWC. 7+ When evaluating model performance, the aforementioned data-driven feature analysis and feature selection are performed only within the training subset of the corresponding cross-validation rounds. Validation samples and test samples do not participate in correlation calculation, mutual information calculation, recursive feature elimination, or embedded feature importance calculation.

[0041] Wherein, CO2g, N2g, H2Sg, C1g, and C2g~C6g represent the mole fractions of CO2, N2, H2S, CH4, and C2~C6 in the injected gas, respectively; C2o~C4o and C5o~C6o represent the mole fractions of C2~C4 and C5~C6 in the crude oil, respectively; MWC 7+ Indicates crude oil C 7+Average molecular weight; Tr represents reservoir temperature; DAG and Vol are determined according to the following formulas: (3) (4) Wherein, H2So, CO2o, Clo, and N2o represent the mole fractions of H2S, CO2, CH4, and N2 in crude oil, respectively. DAG is used to characterize the combined effect of polar dissolved gases in crude oil, and Vol is used to characterize the combined effect of volatile light components in crude oil.

[0042] The necessary parameters of the samples are standardized. The training, validation, and test samples use the same standardized parameters determined by the training samples to avoid information from the validation or test samples being included in the model training process.

[0043] S203. Based on the necessary parameters of the sample and the pre-fitted empirical minimum miscibility pressure model, determine the empirical minimum miscibility pressure of the sample and determine the physical safety range.

[0044] This embodiment uses an empirical model as a provider of physical scope, rather than directly using the calculation results of the empirical model as the final output of the neural network or as a forced training label. The empirical minimum miscibility pressure model can be established based on reservoir temperature, molecular weight of heavy components in crude oil, relative content of volatile and intermediate components in crude oil, and overall critical properties of the injected gas.

[0045] The overall critical properties of the injected gas are determined by the pseudo-critical temperature of the mixed injected gas. Characterization, Determined according to the following formula: (5) in, For the first gas injected The mole fraction of each component; K represents the critical temperature of the corresponding component. This indicates the number of components contained in the sample gas.

[0046] The general formula for the empirical minimum miscibility pressure model is as follows: (6) in, Indicates the empirical minimum miscibility pressure; It represents the sum of the molar percentages of N2 and CH4 in crude oil; It represents the sum of the molar percentages of CO2, H2S, and C2~C6 in crude oil; The value represents the CO2 critical temperature; α, β, λ, δ, and γ represent parameters obtained by fitting sample data. Mixture correction term. Used to characterize the combined effect of changes in injected gas composition on miscibility.

[0047] In a specific fitting result, the empirical minimum miscibility pressure model is: (7) The specific values ​​in formula (7) are example parameters obtained based on the samples in this embodiment. In different oil regions, different experimental sources, or different component ranges, the corresponding samples can be used for refitting, without being limited by the specific parameters in formula (7). When using cross-validation to evaluate the model performance, the parameters of the empirical minimum miscibility pressure model are refitted using the corresponding training subsets in each round of cross-validation, and the fitted parameters are fixed and applied to the corresponding validation subset or test subset; the measured minimum miscibility pressures in the validation subset and test subset do not participate in the fitting of the empirical model parameters.

[0048] Based on experience, minimum miscibility pressure And the deviation coefficient ε, to determine the physical safety range : (8) The allowable deviation coefficient ε can be determined based on the error distribution of the empirical model on independent validation samples or the target coverage. In this embodiment, ε is set to 0.40, corresponding to a validation sample coverage of 91.7%. Figure 3 As shown, predicted values ​​within the physical safety range do not incur empirical range penalties; only values ​​exceeding the upper bound or falling below the lower bound are included in the loss function.

[0049] S204. Construct an automatically differentiable minimum miscibility pressure prediction network to obtain the sample predicted minimum miscibility pressure. For non-component-type parameters such as reservoir temperature and molecular weight, which are not limited by the mole fraction closure condition, calculate the ordinary partial derivative of the sample predicted minimum miscibility pressure with respect to the corresponding parameter; for the injection gas mole fraction components, construct a closed composition perturbation that keeps the sum of the mole fractions of each component of the injection gas at 1, and calculate the directional derivative of the predicted minimum miscibility pressure along the direction of the closed composition perturbation.

[0050] like Figure 2 As shown, the minimum mixing pressure prediction network can employ a fully connected neural network with residual connections. To stably compute the ordinary partial derivatives of uncomponent morphological features and the first derivatives required to construct the closed-form directional derivatives of morphological features, continuously differentiable nonlinear activation functions are used in the hidden layers. The number of hidden layer neurons in a specific network is 256, 128, 64, and 32, respectively, using the Tanh activation function, a dropout ratio of 0.20, and a weight decay of 0.0001.

[0051] Based on reservoir engineering mechanisms and sample response analysis, the minimum miscibility pressure is determined as a function of Tr and MWC. 7+The minimum miscibility pressure increases with increasing N2g and C1g, and decreases with increasing CO2g and H2Sg, provided that the sum of the mole fractions of the injected gas components remains at 1 and the change in the target component is compensated by other injected gas components. This monotonic relationship indicates the direction of the response of the target component relative to the increase of the compensating component, rather than the ordinary partial derivative relationship under the condition that all other components remain constant.

[0052] For the injected gas, the first The target component is selected to increase its mole fraction. And reduce other components according to their relative proportions in the non-target components, so as to keep the sum of the mole fractions of all injected gas components at 1. The mole fractions of the target component and other components after the closed perturbation are shown in equation (9). This represents the original mole fraction of the target component. Indicates the first The original mole fraction of the non-target components, This represents the mole fraction perturbation of the target component and satisfies... According to the above closed perturbation method, the sum of the mole fractions of each injected gas component after perturbation is still 1.

[0053] (9) Sample prediction of minimum miscibility pressure along the first The directional derivative of the closed perturbation direction of the injected gas component is determined according to the following formula (10), where, Indicates the first The directional derivative of the closed-loop composition of each component; This represents the injected gas composition vector after closed perturbation; This represents the predicted minimum miscibility pressure for the sample. When... Less than or equal to the preset boundary threshold At that time, no first step is applied to the sample. Gradient constraints are added to the direction of component addition to avoid numerical instability under pure or near-pure component conditions.

[0054] (10) To unify the representation of the ordinary partial derivative response of uncomponent-type features and the closed-form component directional derivative response of component-type features, the first definition is... Uniform gradient response quantity with pre-defined monotonic features As shown in equation (11). Wherein, Indicates the first A uniform gradient response quantity corresponding to a pre-defined monotonic feature; This represents a set of non-component-type features that are not restricted by the mole fraction closure condition. In this embodiment, it includes reservoir temperature Tr and crude oil C. 7+Average molecular weight (MWC) 7+ ; This represents the set of injection gas component types, expressed as mole fractions and subject to monotonic constraints. In this embodiment, it includes C1g, N2g, CO2g, and H2Sg. For non-component types, The ordinary partial derivative of the predicted minimum miscibility pressure with respect to the corresponding feature is used; for component type features, The closed-form directional derivative is used as defined in the formula. For the first The training samples, and the corresponding uniform gradient response, are denoted as... .

[0055] (11) For the preset positive monotonic features and preset negative monotonic features, the gradient prior loss is determined according to the unified gradient response as shown in equation (12), where, This indicates the number of training samples involved in the gradient constraint calculation; Indicates the first Gradient constraint weights for monotonic features; Represents the set of positive monotonic features, including Tr and MWC. 7+ C1g and N2g; This represents the set of negative monotonic features, including CO2g and H2Sg. When the uniform gradient response conforms to the preset monotonic direction, the corresponding penalty is zero; when the uniform gradient response violates the preset monotonic direction, the penalty increases with the magnitude of the violation.

[0056] (12) Figure 4 This describes the calculation process for monotonic gradient constraints. Input features are standardized according to parameters determined by the training set. Component-specific features also require closure processing while maintaining the sum of the mole fractions of each component at 1. The processed features are input into a physical information neural network, which calculates the predicted minimum miscibility pressure using forward propagation. The uniform gradient response of each preset monotonic feature is obtained through automatic differentiation. The monotonic features of non-component types correspond to the ordinary partial derivatives of the predicted minimum miscibility pressure with respect to the corresponding features. The monotonic characteristic of component types corresponds to the closed-form component directional derivative. When the default is a positive monotonic relationship, the negative part of the uniform gradient response constitutes a single-term gradient penalty, i.e., it adopts... When the default is a negative monotonic relationship, its positive value portion constitutes a one-term gradient penalty, i.e., it adopts... Each individual gradient penalty is applied according to its corresponding gradient constraint weight. Weighted summation to form gradient prior loss This allows the model to impose constraints only on gradient responses that violate the pre-defined physical monotonic direction.

[0057] S205 constructs a joint objective function based on data loss, empirical interval loss, and gradient prior loss, and uses a staged dynamic weighting to train a minimum mixing pressure prediction network. Data loss Experience interval loss and joint objective function Determine them according to the following formulas respectively: (13) (14) (15) in, Indicates the first The measured minimum miscibility pressure of each sample; This indicates the number of samples in a training batch. , and These represent the weights of the data loss, empirical interval loss, and gradient prior loss, respectively.

[0058] like Figure 5 As shown, in a specific implementation with 2000 training rounds, rounds 0 to 600 are the data-driven phase. , , Rounds 601 to 1400 are the phase of progressively increasing physical constraints. It increases linearly from 0.3 to 3.0. It increases linearly from 0.02 to 0.30; rounds 1401 to 2000 are the physically constrained stable phase. , .

[0059] In the specific embodiment described above, a gradient optimization algorithm with a batch size of 64 and an initial learning rate of 0.0015 is used to train the network. The round boundaries, weight start and end values, and weight change functions of the three stages can be adjusted according to the validation samples, or can be adopted using a smooth step, cosine growth, or adaptive changes based on the physical violation rate, as long as the physical constraints are weaker in the early stage of training than in the later stage.

[0060] The parameters updated during joint training are the network parameters of the minimum mixing pressure prediction network. A pre-fitted empirical minimum mixing pressure model is used to determine the physical safety interval, and a preset physical monotonic direction is used to determine the gradient violation. These two are used to form training constraints and are not directly used as the output model of the target minimum mixing pressure.

[0061] S206. Evaluate the trained minimum mixing pressure prediction network on the validation set. Determine the final network parameters based on prediction error, physical safety interval coverage, and monotonic direction consistency. Save the minimum mixing pressure prediction network with determined network parameters and its corresponding standardized parameters as the target minimum mixing pressure prediction model. The target minimum mixing pressure prediction model includes at least network structure information, the network parameters determined after training, and the standardized parameters corresponding to the network input. The test set is only used for independent evaluation of the final model and does not participate in parameter fitting or model selection.

[0062] In conclusion, Figure 2 The offline construction process shown includes: S201, acquiring and cleaning historical miscible experimental samples, and establishing sample group identifiers according to literature sources, oil samples, or experimental series; S202, dividing training samples, validation samples, and test samples according to the sample group identifiers; S203, determining the necessary parameters, fitting standardized parameters, and empirical minimum miscibility pressure model parameters using only training samples, and determining the allowable deviation coefficients and model hyperparameters using independent validation samples; S204, constructing an automatically differentiable residual fully connected neural network and monotonic gradient relationships; S205, training the network based on the joint objective function and dynamic weights; S206, completing model selection based on validation samples, and independently evaluating the final model using only test samples.

[0063] To verify the combined effect of interval constraints, gradient constraints, and dynamic weights, a standard deep neural network was set up as an isomorphic control. The standard deep neural network and the physical information neural network of this invention use the same input parameters, hidden layer size, residual connections, activation functions, batch size, learning rate, and training epochs. The only difference between the two is that the standard deep neural network does not include empirical interval loss, gradient prior loss, and dynamic physical weights.

[0064] This embodiment employs nested five-fold cross-validation to evaluate model performance. Based on the literature source, oil sample, or experimental series corresponding sample group identifiers, all samples are divided into five independent outer folds, ensuring that samples from the same sample group only enter the same outer fold. In each round of outer fold evaluation, one outer fold is selected as the outer test set, and the remaining four outer folds are used as the outer training set. Five-fold cross-validation is further implemented within the outer training set, ensuring that samples from the same sample group only enter the same inner fold. One inner fold is sequentially used as the inner validation set, and the remaining inner folds as the inner training set. Model hyperparameters are determined based on the prediction errors of each inner validation fold. Subsequently, the model is retrained on the corresponding outer training set using the determined model hyperparameters, and independently evaluated on the outer test set (which did not participate in feature selection, standardized parameter determination, empirical minimum miscibility pressure model fitting, allowable deviation coefficient determination, or model selection). The evaluation results of the five outer test folds are summarized, as shown in Table 1. Table 1 As shown in Table 1, the average R² of the physical information neural network of this invention on the five outer test layers is 0.9507±0.0127, the average RMSE is 1.7180±0.5302 MPa, the average MAE is 1.0353±0.1709 MPa, and the average MRE is 0.0574±0.0061. The corresponding evaluation results for the isomorphic deep neural network are 0.9321±0.0127, 2.1281±0.1897 MPa, 1.1170±0.0907 MPa, and 0.0600±0.0067, respectively. Compared with the isomorphic deep neural network, the average RMSE of the physical information neural network of this invention is reduced by 19.27%. In each round of outer layer evaluation, both networks use the same outer training set, outer test set, and inner cross-validation partitioning, and have the same input parameters, network topology, hidden layer size, residual connections, activation functions, batch size, learning rate, and training epochs. The only difference is that the isomorphic deep neural network does not include empirical interval loss, gradient prior loss, and dynamic physical weights. Therefore, the above comparison can reflect the feasibility of combining empirical interval constraints, gradient constraints, and dynamic weights. Figure 6 The table shows the predicted distribution of one of the outer test folds. The values ​​in the table are the mean and standard deviation of the evaluation results of the five outer test folds. These values ​​are used to show the predicted distribution of a single outer test fold and the overall evaluation results of all outer test folds, respectively.

[0065] The target minimum miscibility pressure can be used to determine the pressure search range for capillary experiments, screen different injection gas ratios, set the initial pressure range for component simulation, and evaluate the feasibility of achieving miscibility under reservoir conditions. For target necessary parameters that significantly exceed the applicable range of the training samples, suggestions for supplementary experiments or model refitting can be output.

[0066] Corresponding to the aforementioned embodiments of the crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network, this application also provides embodiments of a crude oil miscibility pressure prediction system using a fusion gradient-constrained physical information network.

[0067] Figure 7 This is a schematic diagram illustrating the structural principle of a crude oil miscibility pressure prediction system that integrates gradient-constrained physical information networks, according to an exemplary embodiment. (Refer to...) Figure 7 The system may include: a parameter acquisition module, a parameter analysis module, a physical interval construction module, a prediction model construction module, a gradient constraint module, a joint training module, and a mixture pressure prediction module; wherein: The parameter acquisition module is used to acquire a training sample set formed by historical miscibility experiments during the model building stage, and to acquire the target crude oil, target gas and target reservoir temperature with unknown minimum miscibility pressure during the target prediction stage. The training sample set includes the measured minimum miscibility pressure as a supervision label, and the measured minimum miscibility pressure of the target crude oil does not need to be acquired during the target prediction stage. The parameter analysis module is used to perform component analysis and feature processing on sample crude oil and sample gas during the model building stage, determine the necessary parameters of the sample and the corresponding standardized parameters, and determine the necessary parameters of the target in the target prediction stage according to the same feature definition as in the model building stage.

[0068] The physical range construction module is used to determine the empirical minimum miscibility pressure and the physical safety range based on the empirical minimum miscibility pressure model. The prediction model construction module is used to construct a minimum miscibility pressure prediction network that takes standardized sample necessary parameters as input and the sample predicted minimum miscibility pressure as scalar output. The gradient constraint module is used to calculate the ordinary partial derivative of the predicted minimum miscibility pressure for non-component monotonic features, calculate the closed composition direction derivative of the component monotonic features to keep the sum of the mole fractions of each component in the corresponding composition system equal to 1, and determine the gradient prior loss based on the difference between one of the ordinary partial derivative and the closed composition direction derivative and the preset physical monotonic direction. The joint training module is used to construct a joint objective function based on data loss, empirical interval loss and gradient prior loss, and update the network parameters of the minimum mixing pressure prediction network using the joint objective function in at least two training stages according to the physical constraint weights from weak to strong. The minimum mixing pressure prediction network with determined network parameters and its corresponding standardized parameters are then saved as the target minimum mixing pressure prediction model. The miscibility pressure prediction module is used to determine the target necessary parameters according to the feature definition adopted in the model construction stage, to standardize the target necessary parameters using the standardized parameters corresponding to the target minimum miscibility pressure prediction model, to input the standardized target necessary parameters into the target minimum miscibility pressure prediction model for forward calculation, and to determine the predicted minimum miscibility pressure output by the target minimum miscibility pressure prediction model as the target minimum miscibility pressure.

[0069] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0070] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0071] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the crude oil miscibility pressure prediction method using a fused gradient-constrained physical information network as described above. Figure 8 The diagram shown is a hardware structure diagram of any device with data processing capabilities in which a gradient-constrained physical information network crude oil miscibility pressure prediction system is provided according to an embodiment of the present invention. Except for... Figure 8 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0072] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the crude oil miscibility pressure prediction method using a fused gradient-constrained physical information network as described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0073] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0074] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting crude oil miscibility pressure using a gradient-constrained physical information network, characterized in that, Includes the model building phase and the target prediction phase: The model building phase includes: Obtain a sample set formed by historical miscibility experiments. Each sample includes sample crude oil, sample gas used to inject the sample crude oil for oil displacement, sample reservoir temperature, and measured minimum miscibility pressure as a monitoring label. The sample crude oil and the sample gas are subjected to component analysis and feature processing to determine the necessary parameters for predicting the minimum miscibility pressure, and the necessary parameters are standardized based on the standardized parameters determined by the training samples. Based on the necessary parameters of the sample and the pre-fitted empirical minimum miscibility pressure model, the empirical minimum miscibility pressure and physical safety range of the sample are determined; a minimum miscibility pressure prediction network with standardized necessary parameters of the sample as input and predicted minimum miscibility pressure of the sample as output is constructed. For non-component-type monotonic features, calculate the ordinary partial derivative of the sample's predicted minimum miscibility pressure with respect to the corresponding feature; for component-type monotonic features expressed in mole fraction, construct a closed compositional perturbation while keeping the sum of the mole fractions of each component in the same compositional system constant, and calculate the closed compositional directional derivative of the sample's predicted minimum miscibility pressure along the direction of the closed compositional perturbation; wherein, monotonic features include non-component-type monotonic features and component-type monotonic features; for reservoir temperature Tr and crude oil C 7+ Average molecular weight (MWC) 7+ Calculate the partial derivative of the predicted minimum miscibility pressure with respect to the corresponding characteristic; for the monotonic characteristic of the injected gas component type, while keeping the sum of the mole fractions of each component of the injected gas at 1, increase the mole fraction of the target component and correspondingly decrease the mole fraction of at least one other injected gas component, thereby constructing a closed composition perturbation direction and calculating the derivative of the closed composition direction; the injected gas N2 content N 2g and CH4 content C 1g The corresponding preset physical monotonic direction is positive, and the injected gas CO2 content is CO 2g The preset physical monotonic direction corresponding to the H2S content H2Sg is negative; the ordinary partial derivatives of non-component monotonic features and the closed component directional derivatives of component monotonic features are uniformly defined as the unified gradient response quantity; for positive monotonic relations, the square of the negative part of the unified gradient response quantity is used to determine the individual gradient penalty, and for negative monotonic relations, the square of the positive part of the unified gradient response quantity is used to determine the individual gradient penalty. The data loss is determined based on the sample predicted minimum miscibility pressure and the measured minimum miscibility pressure. The empirical interval loss is determined based on the extent to which the sample predicted minimum miscibility pressure exceeds the physical safety range. The gradient prior loss is determined based on the difference between one of the ordinary partial derivatives and the derivatives of the closed component direction and the preset physical monotonic direction. A joint objective function is constructed based on the data loss, the empirical interval loss, and the gradient prior loss. The minimum mixing pressure prediction network is trained in at least two training phases with physical constraint weights increasing from weak to strong to obtain the trained minimum mixing pressure prediction network. The trained minimum mixing pressure prediction network and its corresponding standardized parameters are saved, and the trained minimum mixing pressure prediction network is determined as the target minimum mixing pressure prediction model. The target prediction stage includes: The target crude oil with unknown minimum miscibility pressure, the target gas used to inject the target crude oil for oil displacement, and the target reservoir temperature are obtained. The target crude oil and the target gas are subjected to component analysis and feature processing. The necessary parameters of the target are determined according to the same feature definition as in the model building stage. The necessary parameters of the target are then standardized using the standardized parameters saved in the model building stage. The standardized target necessary parameters are input into the target minimum miscibility pressure prediction model for forward calculation, and the predicted minimum miscibility pressure output by the target minimum miscibility pressure prediction model is determined as the target minimum miscibility pressure for the target crude oil and the target gas to mix.

2. The crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network as described in claim 1, characterized in that, The construction process of the target minimum mixing pressure prediction model includes data prediction processing, empirical interval constraint processing, gradient constraint processing, and joint training processing; wherein: The data prediction processing includes: inputting the standardized necessary parameters of the sample into the minimum miscibility pressure prediction network for forward calculation, and outputting the sample predicted minimum miscibility pressure by the minimum miscibility pressure prediction network; The empirical interval constraint processing includes: inputting the necessary parameters of the sample into a pre-fitted empirical minimum miscibility pressure model, calculating the empirical minimum miscibility pressure of the sample, and determining the physical safety interval based on the empirical minimum miscibility pressure of the sample and the allowable deviation coefficient; and determining the empirical interval loss based on the extent to which the predicted minimum miscibility pressure of the sample exceeds the physical safety interval. The gradient constraint processing includes: for non-component-type monotonic features, calculating the ordinary partial derivative of the sample's predicted minimum miscibility pressure with respect to the corresponding feature; for component-type monotonic features, calculating the closed-form composition direction derivative of the sample's predicted minimum miscibility pressure while keeping the sum of the mole fractions of each component in the same composition system at 1; determining the monotonic direction violation amount based on the difference between one of the ordinary partial derivative and the closed-form composition direction derivative and the preset physical monotonic direction, and determining the gradient prior loss based on the monotonic direction violation amount; The joint training process includes: constructing a joint objective function based on the data loss, the empirical interval loss, and the gradient prior loss; updating the network parameters of the minimum mixing pressure prediction network using the joint objective function until a preset training stopping condition is met; and saving the minimum mixing pressure prediction network with determined network parameters as the target minimum mixing pressure prediction model. In the target prediction stage, the target minimum miscibility pressure prediction model outputs the target minimum miscibility pressure based on the standardized target necessary parameters; the empirical minimum miscibility pressure, the physical safety range, the ordinary partial derivative, and the closed-form directional derivative are used to evaluate the physical consistency of the prediction results.

3. The crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network as described in claim 2, characterized in that... The expression for the empirical minimum miscibility pressure model is: ; in, The empirical minimum miscibility pressure is used for the sample and is not included as the final output of the target minimum miscibility pressure prediction model. Tr represents the reservoir temperature, and MWC is used for the measurement. 7+ Crude oil C 7+ Average molecular weight, X vol X is the sum of the molar percentages of N2 and CH4 in crude oil. int The sum of the molar percentages of CO2, H2S, and C2~C6 in crude oil, Tc,CO2 is the critical temperature of CO2, and y i T represents the mole fraction of the i-th component in the injected gas. ci Let be the critical temperature of the i-th component, and α, β, λ, δ, and γ be fitting parameters determined by multiple nonlinear regression based on historical miscibility experimental samples, with the measured minimum miscibility pressure as the dependent variable; the upper and lower bounds of the physical safety interval are respectively... and , The allowable deviation coefficient ε is determined based on the coverage rate of the measured minimum miscibility pressure in the independent validation sample falling within the physical safety range, which reaches a preset coverage threshold.

4. The crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network as described in claim 3, characterized in that, The process of constructing the empirical minimum miscibility pressure model includes: obtaining the necessary parameters of the sample; analyzing the influence of reservoir temperature, molecular weight of heavy components in crude oil, ratio of volatile components to intermediate components in crude oil, and pseudocritical temperature of mixed injection gas on the measured minimum miscibility pressure based on the necessary parameters of the sample; constructing an initial empirical minimum miscibility pressure model based on the influence relationship; fitting the parameters of the initial empirical minimum miscibility pressure model based on the training sample, and determining the allowable deviation coefficient based on the coverage of validation samples that are independent of the training sample; test samples are not involved in the parameter fitting, allowable deviation coefficient determination, and model selection process of the empirical minimum miscibility pressure model.

5. The crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network as described in claim 2, characterized in that, Data loss is determined based on the sample-predicted minimum miscibility pressure and the supervision label; The empirical range loss is determined based on the extent to which the predicted minimum miscibility pressure exceeds the physical safety range. The gradient prior loss is determined based on the difference between one of the ordinary partial derivatives and the closed component directional derivatives and the preset physical monotonic direction. A joint objective function is constructed based on data loss, empirical interval loss, and gradient prior loss. The joint objective function is: ; in, The network parameters of the minimum miscibility pressure prediction network; During joint training, the network parameters are updated according to the joint objective function. After training, the minimum mixing pressure prediction network with determined network parameters is saved as the target minimum mixing pressure prediction model. Training the minimum mixing pressure prediction network includes a data-dominated phase, a physical constraint asymptotic enhancement phase, and a physical constraint stabilization phase, maintaining stability during the data-dominated phase. Keeping the first empirical interval loss weight and the first gradient prior loss weight unchanged, they are continuously increased during the asymptotic enhancement phase of physical constraints. and Maintain the increase after the physical constraint stabilizes. and constant.

6. The crude oil miscibility pressure prediction method using a fusion gradient-constrained physical information network as described in claim 5, characterized in that, The necessary parameters for the sample and the necessary parameters for the target are defined using the same features, including the mole fractions of CO2, N2, H2S, CH4, and C2-C6 in the injected gas; the polar dissolved gas complex characteristic (DAG) of the crude oil; the volatile component complex characteristic (Vol) of the crude oil; and the C2-C4 content, C5-C6 content, and C6 content of the crude oil. 7+ Average molecular weight (MWC) 7+ And reservoir temperature Tr; Where DAG is the sum of the molar percentages of H2S and CO2 in crude oil, and Vol is the sum of the molar percentages of CH4 and N2 in crude oil. The minimum miscibility pressure prediction network is a residual fully connected neural network using continuously differentiable activation functions. In the target prediction stage, the necessary parameters of the target are processed according to the feature arrangement order and standardization method adopted in the model construction stage, and the processed necessary parameters of the target are used as the input of the target minimum miscibility pressure prediction model.

7. A crude oil miscibility pressure prediction system integrating gradient-constrained physical information networks, employing the crude oil miscibility pressure prediction method integrating gradient-constrained physical information networks as described in claim 1, characterized in that... It includes a parameter acquisition module, a parameter analysis module, a physical interval construction module, a prediction model construction module, a gradient constraint module, a joint training module, and a mixed-phase pressure prediction module; among which: The parameter acquisition module is used to acquire a training sample set formed by historical miscibility experiments during the model building stage, and to acquire the target crude oil, target gas and target reservoir temperature with unknown minimum miscibility pressure during the target prediction stage. The training sample set includes the measured minimum miscibility pressure as a supervision label, and the measured minimum miscibility pressure of the target crude oil does not need to be acquired during the target prediction stage. The parameter analysis module is used to perform component analysis and feature processing on sample crude oil and sample gas during the model building stage, determine the necessary parameters of the sample and the corresponding standardized parameters, and determine the necessary parameters of the target in the target prediction stage according to the same feature definition as the model building stage. The physical range construction module is used to determine the empirical minimum miscibility pressure and the physical safety range based on the empirical minimum miscibility pressure model. The prediction model construction module is used to construct a minimum miscibility pressure prediction network that takes standardized sample necessary parameters as input and the sample predicted minimum miscibility pressure as scalar output. The gradient constraint module is used to calculate the ordinary partial derivative of the predicted minimum miscibility pressure for non-component monotonic features, calculate the closed composition direction derivative of the component monotonic features to keep the sum of the mole fractions of each component in the corresponding composition system equal to 1, and determine the gradient prior loss based on the difference between one of the ordinary partial derivative and the closed composition direction derivative and the preset physical monotonic direction. The joint training module is used to construct a joint objective function based on data loss, empirical interval loss and gradient prior loss, and update the network parameters of the minimum mixing pressure prediction network using the joint objective function in at least two training stages according to the physical constraint weights from weak to strong. The minimum mixing pressure prediction network with determined network parameters and its corresponding standardized parameters are then saved as the target minimum mixing pressure prediction model. The miscibility pressure prediction module is used to determine the target necessary parameters according to the feature definition adopted in the model construction stage, to standardize the target necessary parameters using the standardized parameters corresponding to the target minimum miscibility pressure prediction model, to input the standardized target necessary parameters into the target minimum miscibility pressure prediction model for forward calculation, and to determine the predicted minimum miscibility pressure output by the target minimum miscibility pressure prediction model as the target minimum miscibility pressure.

8. A crude oil miscibility pressure prediction device integrating gradient-constrained physical information networks, characterized in that, The system includes a processor, a user interface, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the crude oil miscibility pressure prediction method using a fused gradient-constrained physical information network as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the crude oil miscibility pressure prediction method of the fused gradient-constrained physical information network as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Prediction method for non-pure carbon dioxide miscible flooding minimum miscible pressure GWO-LSSVM

    CN110059435A

  • Asynchronous cycle-based oil extraction apparatus, method and system

    US20260168365A1