Carbon dioxide oil displacement burying multi-objective optimization method based on self-adaptive agent model

Abstract

The invention discloses a carbon dioxide oil displacement burying multi-objective optimization method based on a self-adaptive agent model, and relates to the technical field of oil reservoir injection and production optimization. The method comprises the following steps: firstly, constructing a carbon dioxide flooding embedding injection-production optimization model, then acquiring a plurality of initial samples by adopting Latin hypercube sampling to construct a database, preferably selecting each target agent model based on the database, and then generating a Pareto leading edge by utilizing a dominating class search strategy, a decomposing class search strategy and an index class search strategy; and in the optimization stage, a preferred potential solution is searched according to a hypervolume improvement maximum strategy, numerical simulation is carried out, the database is updated until a preset number of times is reached, each iteration optimization scheme, an oil reservoir net present value and a carbon dioxide burying amount are output, and multi-target optimization of carbon dioxide flooding burying is completed. According to the method, the carbon dioxide flooding multi-objective optimization efficiency is improved, meanwhile, the search direction is dynamically adjusted through the hyper-volume evaluation index, the Pareto frontier is accelerated, and accurate prediction of the carbon dioxide flooding development scheme is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of oil reservoir injection and production optimization, and in particular to a multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model. Background Art

[0002] Optimizing CO2 flooding and storage injection-production schemes aims to design the optimal injection-production scheme based on specific reservoir conditions, thereby improving oil recovery or storage efficiency. However, CO2 flooding and storage projects are characterized by complex physical and chemical reactions, strong reservoir nonlinearity, numerous optimization parameters, and time-consuming numerical simulations. Furthermore, different injection-production schemes vary significantly in oil recovery efficiency, carbon storage capacity, and economic benefits, making it difficult to efficiently obtain an optimized scheme using conventional methods. Therefore, there is an urgent need to explore methods for efficiently optimizing CO2 flooding and storage injection-production schemes.

[0003] Data-driven proxy models map production parameters to reservoir net present value and CO2 storage, using historical production data as input and corresponding economic benefit responses as output for model training. This model can replace traditional numerical simulators, rapidly evaluating candidate development scenarios while significantly reducing the number of numerical simulations. It demonstrates promising application potential and engineering value in reservoir injection and production optimization, and has become a key technical tool for improving optimization efficiency.

[0004] Therefore, it is urgent to apply data-driven proxy models to the formulation of carbon dioxide flooding and storage injection and production plans. A multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive proxy model is proposed to efficiently optimize carbon dioxide flooding and storage injection and production plans. Summary of the Invention

[0005] The present invention aims to overcome the problems of poor adaptability of surrogate models and difficulty in multi-objective coordination in the current CO2 flooding, storage, injection and production optimization methods. A multi-objective optimization method for CO2 flooding, storage and production based on an adaptive surrogate model is proposed. The method improves the robustness and generalization of the surrogate model, enhances its ability to solve high-dimensional complex engineering problems, and combines high-precision prediction accuracy with efficient trade-offs in solution set convergence and diversity. It accelerates the advancement of the Pareto front while achieving accurate prediction of the optimal CO2 flooding solution set, which is conducive to guiding the efficient design of CO2 flooding, injection and production schemes in complex oil reservoirs.

[0006] To achieve the above object, the present invention adopts the following technical solutions: The multi-objective optimization method for CO2 flooding and storage based on an adaptive surrogate model includes the following steps: Step 1: Based on the oilfield geological information and oilfield production information, a CO2 flooding component model is constructed and imported into a reservoir numerical simulator; Step 2: Build a CO2 flooding storage and injection-production optimization model. Use Latin hypercube sampling to obtain multiple injection-production scenarios to construct an initial sample set. Use a reservoir numerical simulator to calculate the net present value and CO2 storage capacity corresponding to each sample in the initial sample set, and then build a database. Step 3: Divide the samples in the database into training sets and test sets, and build proxy models for each optimization objective based on the samples in the training set; Step 4: Use the samples in the test set to obtain the root mean square error of the proxy model corresponding to each optimization target. Based on the determination coefficient, select the proxy model with the best prediction accuracy for each optimization target in the current optimization stage, and adapt the proxy model of each optimization target in the current optimization stage. Step 5: Generate offspring populations using genetic algorithms and evaluate offspring individuals based on the proxy models of each optimization objective; Step 6: Construct a hypervolume-guided search strategy selection mechanism to select potential solutions based on the hypervolume evaluation index. Step 7: Repeat steps 3 to 6 to perform numerical simulation on the optimal potential solution until the preset total number of evaluations MaxFEs is reached, and the updated database is obtained. ; Step 8: After the iterative optimization is completed, the reservoir numerical simulator outputs the samples of each iterative optimization and the corresponding reservoir net present value and CO2 storage capacity, obtains the Pareto front solution and the corresponding target results, and completes the multi-objective optimization of CO2 flooding and storage.

[0007] Preferably, in step 1, a three-dimensional geological model is constructed based on oilfield geological data using oilfield geological modeling software, and the three-dimensional geological model is imported into a reservoir numerical simulator. Then, based on oilfield production information, numerical simulation software is used to simulate and obtain fluid high-pressure physical properties and oil-gas-water relative permeability curves, and a fluid model is determined. The three-dimensional geological model and the fluid model are combined to construct a carbon dioxide flooding component model, and the carbon dioxide flooding component model is imported into the reservoir numerical simulator.

[0008] Preferably, step 2 includes the following sub-steps: Step 2.1: For the optimization problem of CO2 flooding and storage and injection-production, a CO2 flooding net present value model is constructed with maximizing the net present value of the reservoir and maximizing the CO2 storage volume as the dual objective function, and the development and control parameters of production wells and gas injection wells as the decision variables, combined with the variable boundary constraints. The net present value model of carbon dioxide flooding is: ; The constraints are: ; Where, is the economic net present value; is the time step number; is the reservoir production time; For the time steps; is the asset discount rate; is the well number; is the number of producing wells; For the Time step The oil production rate of the production wells; is the crude oil price; For the Time step Gas production rate of production wells; To process the produced gas cost; For the Time step The water production rate of the production well; is the produced fluid treatment cost; is the number of injection wells; For the Time step Gas injection rate of the injection well; is the gas injection cost; is the bottom hole flowing pressure of the production well; It is the lower limit of bottom hole flowing pressure of production well; The upper limit of the bottom hole flowing pressure of the production well; is the gas injection rate of the injection well; is the lower limit of the gas injection rate of the injection well; is the upper limit of the gas injection rate of the injection well; Step 2.2, based on the change in carbon dioxide concentration within the grid of the carbon dioxide flooding component model in the reservoir numerical simulator before and after injection and production, calculate the total carbon dioxide storage capacity and construct a carbon dioxide flooding storage capacity model; The carbon dioxide landfill inventory model is: ; Where, is the amount of carbon dioxide stored; is the grid number; is the total number of grids; For the The volume of the grid; Before injection The molar concentration of carbon dioxide per grid; After injection and production The molar concentration of carbon dioxide per grid; is the molar mass of carbon dioxide; Step 2.3: Combine the CO2 flooding net present value model and the CO2 flooding storage model to establish a CO2 flooding oil storage injection and production optimization model and import it into the reservoir numerical simulator; Step 2.4: Use Latin hypercube sampling to obtain multiple samples within the decision variable constraint domain and construct the initial sample set. , ,in, For the The injection plan of each sample, is the initial sample size; Step 2.5: Import the initial sample set into the reservoir numerical simulator, and use the reservoir numerical simulator to calculate the net present value of the reservoir and the carbon dioxide storage capacity of each sample, and obtain the initial sample set. Corresponding dual-target dataset ,in, is the reservoir net present value dataset, , For the The net present value of the reservoir of the sample, is the carbon dioxide storage data set, , For the The amount of CO2 stored in each sample; Step 2.6, based on the initial sample set and its corresponding reservoir net present value dataset and CO2 storage datasets , build the database , .

[0009] Preferably, step 3 includes the following sub-steps: Step 3.1: Get the total sample size of the database , randomly selected The samples are used to form a training set, and the remaining samples are used to form a test set; the training set includes a reservoir net present value training set and CO2 storage training set , where the reservoir net present value training set , CO2 storage training set ; Step 3.2: constructing a proxy model based on the samples in the training set, wherein the proxy model includes a radial basis function proxy model and a Gaussian process regression model. The radial basis function proxy model is constructed based on a Gaussian kernel function and an inverse multi-quadratic kernel function, and the Gaussian process regression model is constructed based on an autocorrelation squared exponential kernel function. The radial basis function proxy model is: ; in, When the kernel function is a Gaussian kernel function, the expression of the kernel function is: ; When the kernel function is an inverse multi-quadratic kernel function, the expression of the kernel function is: ; Where, is the radial basis function; is the Euclidean distance between the input point and the center point; is the serial number of the kernel function; is the total number of kernel functions; For the The weight of the kernel function; is the kernel function; is the bias term; is the shape parameter used to determine the function expansion; is the shape parameter used to determine the smoothness of the function; The Gaussian process regression model is: ; in, The autocorrelation squared exponential kernel function for: ; Where, 、 All are input points; is the fitting function; is the Gaussian process function; is the mean of the distribution; is the kernel function used to measure the similarity between any input points; To calculate the expected value of a random variable; is the signal variance; is an exponential function; is the dimension number; is the input dimension; is the length scale of the input dimension; For input point No. The value of the dimensional input variable; For input point No. The value of the input variable.

[0010] Preferably, step 4 includes the following sub-steps: Step 4.1: Substitute the reservoir net present value and CO2 storage capacity of each test sample in the test set into a radial basis function proxy model with a Gaussian kernel, an inverse multi-quadratic kernel function, and a Gaussian process regression model with an autocorrelation squared exponential kernel function, respectively, and determine the coefficient of determination corresponding to each proxy model; The determination coefficient calculation formula is: ; Where, is the coefficient of determination; is the test sample serial number; is the true value of the surrogate model; is the predicted value of the surrogate model; is the average value of the true value of the test sample; In step 4.2, the proxy model with the determination coefficient closest to 1 among the proxy models is selected as the best proxy model for the optimization target, and the proxy model with the best prediction accuracy for each optimization target in the current optimization stage is obtained, and the proxy model of each optimization target in the current optimization stage is adapted to suit the needs of the optimization target.

[0011] Preferably, step 5 includes the following sub-steps: Step 5.1, set the number of updates and potential samples of the adaptive agent model stage, set the total number of reservoir numerical simulation evaluations MaxFEs, and use the genetic algorithm to generate the offspring population; The process of generating offspring population by the genetic algorithm includes gene exchange and gene mutation; The gene exchange process is: ; ; The gene mutation process is: ; Where, is the individual serial number, is the gene number, is the total number of genes; The first individual; The parent individual 1 genes; The parent individual 2 genes; Control parameters for the degree of offspring offset; is a random number; is the distribution index for controlling crossover; The first a new individual; The first individual; is the disturbance factor, which follows a multinomial distribution; is the upper limit of the variable; is the lower limit of the variable; In step 5.2, the proxy model of each optimization objective is used to predict the net present value of the oil reservoir and the carbon dioxide storage capacity corresponding to the individuals in the offspring population.

[0012] Preferably, step 6 includes the following sub-steps: Step 6.1, set the number of hypervolume stage updates and determine whether the hypervolume improvement set needs to be updated in stages If an update is required, the dominance search strategy, decomposition search strategy and index search strategy are traversed to optimize the potential solution; The dominance search strategy uses a fast non-dominated sorting genetic algorithm to calculate the population dominance relationship, selectively sorts individuals in the same non-dominated layer based on their crowding distance, and selects individuals with larger crowding distance as potential solutions; The decomposition search strategy decomposes the high-dimensional multi-objective optimization problem into several single-objective sub-problems, enhances the control ability of the objective trade-off relationship through reference points or weight vectors, and guides the evolutionary algorithm to select potential solutions based on the reference vectors and update the reference vectors; The index-based search strategy uses performance indicators to quantitatively evaluate individuals in the population, and balances population diversity and algorithm convergence through fitness value update and environment selection mechanism. By calculating the updated fitness values of all individuals in the population and sorting them, the current worst individual is deleted and the population is updated, and the fitness value of the updated population is calculated until a potential solution is screened out; Step 6.2, perform numerical simulation on each potential solution selected, and initialize the hypervolume improvement set , ,in, The hypervolume increment for the dominant search strategy, For decomposition class search strategy hypervolume increment, Search strategy for indicator class super volume increment; The hypervolume improvement value is used to quantify the potential of each search strategy to advance the Pareto frontier. By calculating the hypervolume difference between the Pareto frontier and the reference point before and after optimization, the improvement of the search strategy on solution convergence and diversity is intuitively evaluated. The hypervolume improvement set is initialized. ,get: ; Where, For the The improvement value of the hypervolume after iterations; For the The set of non-dominated solutions after iterations; is the hypervolume calculation function; is the number of iterations; For the The reference points generated after iterations; Step 6.3, if no stage is required to update the hypervolume improvement set , then select the hypervolume improvement set The search strategy with the largest average value selects the potential solution and performs numerical simulation to update the database. , calculate the current Pareto frontier relative to the previous optimized Pareto frontier super volume improvement value , update the hypervolume improvement set, the updated hypervolume improvement set is: ; in, For the The hypervolume increment of the dominant search strategy after iterations, For the The decomposition search strategy hypervolume increment after iterations, For the The hypervolume increment of the indicator-based search strategy after iterations.

[0013] Preferably, the individual crowding calculation formula in the dominant search strategy is: ; Where, For individuals crowding distance; is the ordinal number of the target dimension; is the total number of target dimensions; For individuals In the target dimension Target value on Target dimension Upper maximum value; Target dimension Upper minimum value; The process of optimizing potential solutions by the decomposition search strategy is as follows: First, the target value is normalized based on the ideal point to obtain: ; Secondly, the population is divided and the individuals are assigned to the subpopulation corresponding to the expected closest reference vector, and the result is: ; Finally, based on the convergence and diversity of the angle-penalized distance balance solution, we obtain: ; Where, For individuals Target value before conversion; For individuals The target value after conversion; For the current population The minimum target value in dimensions; For the current population Target value for each dimension; is the reference vector the corresponding subpopulation; For the individual; For individuals The serial number of the reference vector corresponding to the maximum cosine angle between it and the reference vector; is the complex argument calculation function; is the maximum value function; is the cosine function; is the arccosine function; is the two-norm; is the serial number of the reference vector; is the total number of reference vectors; For individuals With reference vector The cosine angle between For the reference vectors; For individuals With reference vector The cosine angle between For individuals To the reference vector Angle penalty distance; is the penalty function; is the current iteration number; is the maximum number of iterations; is the rate of change control parameter of the penalty function; is the reference vector Minimum angle with other reference vectors; The calculation formula for updating the reference vector in the decomposition search strategy is: ; Where, For the After the adjustment reference vectors; The first reference vectors; is the element-wise product of vectors; For the The maximum value of each objective function in the generation; For the The minimum value of each objective function in the generation; The quantitative evaluation process of the individual population by the indicator search strategy is as follows: ; ; ; Where, is the individual serial number; Calculate the function for the indicator value between two individuals; is the normalized individual The value of the optimization target; To optimize the target sequence number; is the total number of optimization targets; For individuals in The original value of the optimization target; is the minimum function; is the individual fitness value; is the total number of individuals in the population; is the normalized index value; is the scaling factor; is the fitness value of the individual after update.

[0014] Preferably, in step 8, after the iterative optimization is completed, the non-dominated solution set is obtained based on the dominance relationship of the solution. : ; Where, 、 All are Chinese databases individuals in It is a Pareto dominance relationship; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; The Pareto solution set is the set of all Pareto optimal solutions in the decision space. The Pareto solution set constructed using non-dominated solutions is: ; Where, is the target value of all dimensions corresponding to the individual; For decision-making space.

[0015] The beneficial technical effects brought about by the present invention are: (1) The present invention proposes a multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive proxy model. The adaptive proxy model is used to automatically adapt the most suitable proxy model according to the characteristics of the task database at different stages of injection and production optimization. The hysteresis effect of displacement effect, storage volume change, etc. in the process of carbon dioxide flooding and storage is fully considered. The dynamic changes of the reservoir and the injection and production rules can be accurately captured, and the prediction results are more accurate and reliable. The cost of numerical simulation calculation is greatly reduced, and the optimal injection and production plan can be obtained with limited resources, which is conducive to efficient optimization.

[0016] (2) This invention proposes a multi-objective optimization method for CO2 flooding and storage based on an adaptive proxy model. Based on the hypervolume-guided search mechanism, it can flexibly adapt to the dynamic changes of the database, continuously optimize the hypervolume-guided search strategy, and enhance the generalization ability of the multi-objective optimization method for CO2 flooding and storage. At the same time, during the search process, the hypervolume-guided search strategy aims to maximize the hypervolume. By quantifying the pros and cons of the multi-objective optimization results, it systematically guides the search direction and quickly obtains a set of high-quality Pareto front solutions covering different preferences. This provides a rich and scientific decision-making basis for the formulation of CO2 flooding and storage injection and production plans, showing extremely high engineering application value and broad development prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the multi-objective optimization method for carbon dioxide flooding and storage based on the adaptive agent model of the present invention.

[0018] Figure 2 Flowchart for determining the adaptive agent model for the present invention.

[0019] Figure 3 The present invention is a flow chart of optimizing potential solutions based on the hypervolume evaluation index-guided search strategy selection mechanism.

[0020] Figure 4 This is the permeability field map of the study block; in the figure, P1~P9 are production wells, and I1~I4 are injection wells.

[0021] Figure 5 is the Pareto front generated during the optimization process; in the figure, ASHMO is the method of the present invention, KRVEA is the traditional multi-objective optimization method based on the Kriging model, NSGA-Ⅲ is a hybrid multi-objective optimization algorithm, KTA2 is a multi-objective optimization algorithm proposed for expensive optimization problems, and NSGA3 is a traditional single-objective optimization method.

[0022] Figure 6Schematic diagram of the performance of the proxy model in the later stage of optimization; in the figure, (a) is a schematic diagram of the net present value of the oil reservoir, and (b) is a schematic diagram of the carbon dioxide storage capacity.

[0023] Figure 7 This is a schematic diagram of the calling of each search strategy during the optimization process.

[0024] Figure 8 Schematic diagram of the production situation of four typical strategic schemes; in the figure, (a) is the curve of the cumulative oil production of the four typical strategic schemes changing with the production time, (b) is the curve of the cumulative carbon dioxide injection volume of the four typical strategic schemes changing with the production time, (c) is the curve of the cumulative carbon dioxide storage volume of the four typical strategic schemes changing with the production time, and (d) is the curve of the cumulative water production of the four typical strategic schemes changing with the production time.

[0025] Figure 9 The figures are the volume molar concentration diagrams of carbon dioxide for four typical strategic schemes; in the figures, (a) is the volume molar concentration diagram of carbon dioxide for the first typical strategic scheme, (b) is the volume molar concentration diagram of carbon dioxide for the second typical strategic scheme, (c) is the volume molar concentration diagram of carbon dioxide for the third typical strategic scheme, and (d) is the volume molar concentration diagram of carbon dioxide for the fourth typical strategic scheme.

[0026] Figure 10 These are the remaining oil field maps of four typical strategic schemes; in the figure, (a) is the remaining oil field map of the first typical strategic scheme, (b) is the remaining oil field map of the second typical strategic scheme, (c) is the remaining oil field map of the third typical strategic scheme, and (d) is the remaining oil field map of the fourth typical strategic scheme. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0028] Example 1 In this embodiment, a multi-objective optimization method for CO2 flooding and storage based on an adaptive agent model is proposed. Figure 1 As shown, the specific steps include: Step 1: Based on the oilfield geological data, a 3D geological model is constructed using the oilfield geological modeling software Petrol. The 3D geological model is imported into the reservoir numerical simulator. Then, based on the oilfield production information, the numerical simulation software Eclipse is used to simulate the fluid high-pressure physical properties and oil-gas-water relative permeability curves, determine the fluid model, and combine the 3D geological model and fluid model to construct a carbon dioxide flooding component model, which is stored in the reservoir numerical simulator.

[0029] Step 2: Build a CO2 flooding, storage, injection, and production optimization model. Use Latin hypercube sampling to obtain multiple injection and production schemes to construct an initial sample set. Use a reservoir numerical simulator to calculate the net present value and CO2 storage volume corresponding to each sample in the initial sample set, and build a database to provide basic data support for the subsequent construction of the proxy model.

[0030] The step 2 includes the following sub-steps: In step 2.1, for the optimization problem of CO2 flooding and storage and injection-production, a CO2 flooding net present value model is constructed with maximizing the net present value of the reservoir and maximizing the CO2 storage volume as the dual objective function, the development and control parameters of production wells and gas injection wells as the decision variables, and the variable boundary constraints.

[0031] The net present value model of carbon dioxide flooding is: ; The constraints are: ; Where, is the economic net present value, in USD; is the time step number; is the reservoir production time; For the Time step, in days; is the asset discount rate; is the well number; is the number of production wells, in units of wells; For the Time step The oil production rate of the production well, in m³·d -1 ; is the crude oil price, in USD·m -3 ; For the Time step Gas production rate of production wells, in m³·d -1 ; The cost of processing the produced gas, in USD·m -3 ; For the Time step The water production rate of the production well, in m³·d -1 ; is the produced fluid treatment cost, in USD·m -3 ; is the number of injection wells, in units of wells; For the Time step Gas injection rate of the injection well, in m³·d -1 ; is the gas injection cost, in USD·m -3 ; is the bottom hole flowing pressure of the production well; It is the lower limit of bottom hole flowing pressure of production well; The upper limit of the bottom hole flowing pressure of the production well; is the gas injection rate of the injection well; is the lower limit of the gas injection rate of the injection well; It is the upper limit of gas injection rate of injection well.

[0032] Step 2.2: Based on the changes in CO2 concentration within the grid of the CO2 flooding component model in the reservoir numerical simulator before and after injection and production, the total CO2 storage capacity is calculated and a CO2 flooding storage capacity model is constructed.

[0033] The carbon dioxide landfill inventory model is: ; Where, is the carbon dioxide storage capacity, in kg; is the grid number; is the total number of grids; For the The volume of the grid, in m 3 ; Before injection The molar concentration of carbon dioxide in the grid, in mol·m -3 ; After injection and production The molar concentration of carbon dioxide in the grid, in mol·m -3 ; is the molar mass of carbon dioxide, in kg·mol -1 .

[0034] In step 2.3, the CO2 flooding net present value model and the CO2 flooding storage model are combined to establish a CO2 flooding oil storage injection and production optimization model and store it in the reservoir numerical simulator.

[0035] Step 2.4: Use Latin hypercube sampling to obtain multiple samples within the decision variable constraint domain and construct the initial sample set. , ,in, For the The injection plan of each sample, is the initial sample size.

[0036] Step 2.5: Import the initial sample set into the reservoir numerical simulator, and use the reservoir numerical simulator to calculate the net present value of the reservoir and the carbon dioxide storage capacity of each sample, and obtain the initial sample set. Corresponding dual-target dataset ,in, is the reservoir net present value dataset, , For the The net present value of the reservoir of the sample, is the carbon dioxide storage data set, , For the The amount of carbon dioxide stored in each sample.

[0037] Step 2.6, based on the initial sample set and its corresponding reservoir net present value dataset and CO2 storage datasets , build a database containing multidimensional decision variables and dual objective function values , .

[0038] Step 3: Divide the samples in the database into a training set and a test set, and construct a proxy model for each optimization objective based on the samples in the training set. The proxy models in this embodiment include a radial basis function proxy model with a Gaussian kernel function, a radial basis function proxy model with an inverse multi-quadratic kernel function, and a Gaussian process regression model with an auto-correlation square exponential kernel function.

[0039] Step 3 includes the following sub-steps: Step 3.1: Get the total sample size of the database , randomly selected The samples are used to form a training set, and the remaining samples in the database are used to form a test set for model verification; the training set includes a reservoir net present value training set and CO2 storage training set ,in, , .

[0040] Step 3.2, constructing a proxy model based on the samples in the training set, the proxy model includes a radial basis function proxy model and a Gaussian process regression model; the radial basis function proxy model is specifically a radial basis function proxy model with a Gaussian kernel function, and a radial basis function proxy model with an inverse multi-quadratic kernel function, wherein the radial basis function proxy model with a Gaussian kernel function is constructed based on the Gaussian kernel function, and the radial basis function proxy model with an inverse multi-quadratic kernel function is constructed based on the inverse multi-quadratic kernel function; the Gaussian process regression model is specifically a Gaussian process regression model with an auto-correlation square exponential kernel function, which is constructed based on the auto-correlation square exponential kernel function.

[0041] The radial basis function proxy model is: ; in, When the kernel function is a Gaussian kernel function, the expression of the kernel function is: ; When the kernel function is an inverse multi-quadratic kernel function, the expression of the kernel function is: ; Where, is the radial basis function; is the Euclidean distance between the input point and the center point; is the serial number of the kernel function; is the total number of kernel functions; For the The weight of the kernel function; is the kernel function; is the bias term, which is used to adjust the final output value; is the shape parameter used to determine the function expansion; is the shape parameter that determines the smoothness of the function.

[0042] The Gaussian process regression model is: ; in, The autocorrelation squared exponential kernel function for: ; Where, 、 All are input points; is the fitting function; is the Gaussian process function; is the mean of the distribution; is the kernel function used to measure the similarity between any input points; To calculate the expected value of a random variable; is the signal variance, which is used to represent the amplitude of the function value; is an exponential function; is the dimension number; is the input dimension; is the length scale of the input dimension, and different dimensions have different correlations; For input point No. The value of the dimensional input variable; For input point No. The value of the input variable.

[0043] Step 4: Use the samples in the test set to obtain the root mean square error of the proxy model corresponding to each optimization target, select the proxy model with the best prediction accuracy for each optimization target in the current optimization stage based on the determination coefficient, and determine the proxy model for each optimization target in the current optimization stage.

[0044] The step 4 includes the following sub-steps: In step 4.1, the reservoir net present value and CO2 storage capacity of each test sample in the test set are respectively substituted into the radial basis function proxy model with Gaussian kernel and inverse multi-quadratic kernel function and the Gaussian process regression model with autocorrelation square exponential kernel function, and the determination coefficient corresponding to each proxy model is determined.

[0045] The determination coefficient calculation formula is: ; Where, is the coefficient of determination; is the test sample serial number; is the true value of the surrogate model; is the predicted value of the surrogate model; is the average value of the true value of the test sample.

[0046] The coefficient of determination By comparing the total variance between the predicted value of the surrogate model and the true value, we can quantify the fitting effect of the surrogate model on the overall data and determine the coefficient The closer it is to 1, the stronger the explanatory ability of the surrogate model for the response variable, which is used as an indicator to measure the goodness of fit of the surrogate model.

[0047] Step 4.2: Determine the coefficients in the radial basis function surrogate model with Gaussian kernel function, the radial basis function surrogate model with inverse multi-quadratic kernel function, and the Gaussian process regression model with autocorrelation square exponential kernel function. The proxy model closest to 1 is taken as the best proxy model for the optimization target, and the proxy model with the best prediction accuracy for each optimization target in the current optimization stage is obtained. The proxy model for each optimization target in the current optimization stage is determined, such as Figure 2 shown.

[0048] Step 5: Generate offspring populations using genetic algorithms and evaluate the proxy models for each optimization objective.

[0049] The step 5 includes the following sub-steps: Step 5.1: Set the number of updates and potential samples for the adaptive agent model stage, set the total number of reservoir numerical simulation evaluations MaxFEs, and use a genetic algorithm to generate a progeny population carrying the genetic information of both parents using simulated binary crossover, and use polynomial mutation to maintain population diversity. The genetic algorithm generates the progeny population through gene exchange and gene mutation.

[0050] The gene exchange process is: ; ; The gene mutation process is: ; Where, is the individual serial number, is the gene number, is the total number of genes; The first individual; The parent individual 1 genes; The parent individual 2 genes; Control parameters for the degree of offspring offset; It is a random number ranging from 0 to 1, used to adjust the selection of mutation strategy; To control the distribution index of the crossover, The bigger it is, the stronger the global search capability is. The smaller the local development capability, the stronger it is; The first a new individual; The first individual; is the disturbance factor, which follows a multinomial distribution; is the upper limit of the variable; is the lower limit of the variable; and variable lower limits Used to ensure that the mutation results are within the feasible range.

[0051] In step 5.2, the proxy model of each optimization objective is used to predict the net present value of the oil reservoir and the carbon dioxide storage capacity corresponding to the individuals in the offspring population.

[0052] Step 6: Construct a hypervolume-guided search strategy selection mechanism, and select potential solutions based on the hypervolume evaluation index, such as Figure 3 shown.

[0053] The step 6 includes the following sub-steps: Step 6.1, set the number of hypervolume stage updates and determine whether the hypervolume improvement set needs to be updated in stages If an update is required, the dominance search strategy, decomposition search strategy and indicator search strategy are traversed to optimize the potential solution.

[0054] The dominance search strategy adopts the fast non-dominated sorting genetic algorithm (NSGA-Ⅱ) to calculate the population dominance relationship, selectively sorts individuals in the same non-dominated layer based on their crowding distance, and selects individuals with larger crowding distance as potential solutions.

[0055] The individual crowding calculation formula is: ; Where, For individuals crowding distance; is the ordinal number of the target dimension; is the total number of target dimensions; For individuals In the target dimension Target value on Target dimension Upper maximum value; Target dimension Upper minimum value.

[0056] By normalizing each target dimension, the bias caused by differences in target scales can be avoided, effectively improving the uniformity and diversity of population distribution.

[0057] The decomposition search strategy decomposes the high-dimensional multi-objective optimization problem into several single-objective sub-problems, enhances the control ability of the objective trade-off relationship through reference points or weight vectors, and optimizes the potential solution based on the reference vector guided evolutionary algorithm (RVEA) and updates the reference vector. The process of optimizing the potential solution is as follows:

[0058] First, the target value is normalized based on the ideal point to obtain: ; Secondly, the population is divided and the individuals are assigned to the subpopulation corresponding to the expected closest reference vector, and the result is: ; Finally, based on the convergence and diversity of the angle-penalized distance balance solution, we obtain: ; Where, For individuals Target value before conversion; For individuals The target value after conversion; For the current population The minimum target value in each dimension is used to set the minimum value of each target value as the ideal point; For the current population Target value for each dimension; is the reference vector the corresponding subpopulation; For the individual; For individuals The serial number of the reference vector corresponding to the maximum cosine angle between it and the reference vector; is the complex argument calculation function; is the maximum value function; is the cosine function; is the arccosine function; is the two-norm; is the serial number of the reference vector; is the total number of reference vectors; For individuals With reference vector The cosine angle between them, the larger the cosine angle, the closer the target vector is to the reference vector. With vector When the angle between them is the smallest, the individual Divided into ; For the reference vectors; For individuals With reference vector The cosine angle between For individuals To the reference vector Angle penalty distance; is the penalty function; is the current iteration number; is the maximum number of iterations; is the change rate control parameter of the penalty function, and its value range is 1~2; is the reference vector The minimum angle between the vector and any other reference vector.

[0059] The calculation formula for updating the reference vector in the decomposition search strategy is: ; Where, For the After the adjustment reference vectors; The first reference vectors; is the element-wise product of vectors; For the The maximum value of each objective function in the generation; For the The minimum value of each objective function in the generation.

[0060] The index-based search strategy uses performance indicators to quantitatively evaluate population individuals and balances population diversity and algorithm convergence through fitness value update and environment selection mechanism. The quantitative evaluation process is as follows: ; ; ; Where, is the individual serial number; Calculate the function for the indicator value between two individuals; is the normalized individual The value of the optimization target; To optimize the target sequence number; is the total number of optimization targets; For individuals in The original value of the optimization target; is the minimum function; is the individual fitness value; is the total number of individuals in the population; is the normalized index value; is the scaling factor, the default value is 0.05, the scaling factor Directly affects the convergence characteristics and scope of the exponential decay term in the fitness calculation, thereby controlling the sensitivity of the optimization algorithm to performance index differences. The scaling factor has been verified by experiments. When the value is 0.05, the algorithm has better convergence and diversity; It is the fitness value of the updated individual, which combines the current individual fitness and its impact on other individuals.

[0061] By calculating the updated fitness values of all individuals in the population and sorting them, deleting the current worst individual and updating the population, and calculating the fitness value of the updated population, a potential solution is screened out.

[0062] Step 6.2, perform numerical simulation on each potential solution selected, and initialize the hypervolume improvement set .

[0063] The hypervolume improvement value is used to quantify the potential of each search strategy to advance the Pareto frontier. By calculating the hypervolume difference between the Pareto frontier and the reference point before and after optimization, the improvement of the search strategy on solution convergence and diversity is intuitively evaluated. The hypervolume improvement set is initialized. ,get: ; Where, For the The improvement value of the hypervolume after iterations; For the The set of non-dominated solutions after iterations; is the hypervolume calculation function; is the number of iterations; For the The reference points generated after iterations.

[0064] In order to more accurately represent the change in hypervolume before and after optimization, the two hypervolume calculations use the same reference point to initialize the hypervolume improvement set , ,in, The hypervolume increment for the dominant search strategy, For decomposition class search strategy hypervolume increment, Search strategy for indicator class hypervolume increment.

[0065] Step 6.3, if no stage is required to update the hypervolume improvement set , then select the hypervolume improvement set The search strategy with the largest average value selects the potential solution and performs numerical simulation to update the database. , calculate the current Pareto frontier relative to the previous optimized Pareto frontier super volume improvement value , update the hypervolume improvement set, the updated hypervolume improvement set is: ; in, For the The hypervolume increment of the dominant search strategy after iterations, For the The decomposition search strategy hypervolume increment after iterations, For the The hypervolume increment of the indicator-based search strategy after iterations.

[0066] Step 7: Repeat steps 3 to 6 to perform numerical simulation on the optimal potential solution until the preset total number of evaluations MaxFEs is reached, and the updated database is obtained. .

[0067] Step 8: After the iterative optimization is completed, the reservoir numerical simulator outputs the samples of each iterative optimization and the corresponding reservoir net present value and carbon dioxide storage volume, obtains the Pareto front solution and the corresponding target result, and obtains the non-dominated solution of carbon dioxide flooding and the numerical simulation results, namely the injection-production plan, reservoir net present value and carbon dioxide storage volume, thus completing the multi-objective optimization of carbon dioxide flooding and storage.

[0068] Specifically, after the iterative optimization is completed, the non-dominated solution set is obtained based on the dominance relationship of the solution. : ; Where, 、 All are Chinese databases individuals in It is a Pareto dominance relationship; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension.

[0069] Pareto dominance is used as a criterion to judge whether a solution is better than another solution. The Pareto solution set is the set of all Pareto optimal solutions in the decision space. The Pareto solution set constructed using non-dominated solutions is: ; Where, is the target value of all dimensions corresponding to the individual; For decision-making space.

[0070] Example 2 In order to verify the effect of the method of the present invention on the multi-objective optimization of carbon dioxide flooding and storage, this example adopts the multi-objective optimization method of carbon dioxide flooding and storage based on the adaptive agent model described in Example 1 and applies it to a certain research block. The research block adopts a five-point well pattern and uses Eclipse 300 to build a carbon dioxide flooding component model for oil, gas and water three-phase drive. The size of the carbon dioxide flooding component model is , the reservoir porosity is 0.17, the total number of grids is 625, and the grid size is The CO2 flooding component model includes 9 production wells and 4 water injection wells, with a production time of 1830 days. The design is divided into 10 time steps, each time step is 183 days. The gas injection rate and the bottom hole pressure of the production wells are both CO2 flooding optimization variables. Among them, the upper limit of the bottom hole pressure of the production well is 29.00MPa and the lower limit is 19.00MPa. The upper limit of the gas injection rate is 15000m³·d -1 , the lower limit is 5000 m³·d -1 , the total number of variables to be optimized is 130 dimensions.

[0071] This example adopts the multi-objective optimization method for CO2 flooding and storage based on the adaptive agent model described in Example 1, and sets the total number of reservoir numerical simulation evaluations MaxFEs to , total sample size The number of updates in the hypervolume stage is 20, the number of updates in the adaptive proxy model stage is 20, and the evaluation sample size is set to 5, that is, the number of potential solutions is 5. The specific process is: Step 1: Based on the oilfield geological data, a three-dimensional geological model is constructed using the oilfield geological modeling software Petrol. The three-dimensional geological model is imported into the reservoir numerical simulator. Then, based on the oilfield production information, the numerical simulation software Eclipse300 is used to simulate the fluid high-pressure physical properties and oil-gas-water relative permeability curve to determine the fluid model. Figure 4 As shown, the three-dimensional geological model and the fluid model are combined to construct a carbon dioxide flooding component model, and the carbon dioxide flooding component model is imported into the reservoir numerical simulator.

[0072] Step 2: Build a CO2 flooding storage and injection optimization model and determine the upper limit of the CO2 flooding gas injection rate per well to be 15,000 m³·d -1 , the lower limit is 5000m³·d -1 The upper limit of the bottom hole pressure of the production well is 29.00MPa and the lower limit of the bottom hole pressure is 19.00MPa. 100 initial sample sets are obtained by using Latin hypercube sampling, and the reservoir production time is set. The time step is 1830 days. is 183 days, and the asset discount rate 0, the number of production wells 9, crude oil price 460 USD·m -3 ; Cost of processing produced gas 0.027 USD·m -3 , produced fluid treatment cost 31.5 USD·m -3 ; Number of injection wells 4; gas injection cost 0.083 USD·m -3 Then, the net present value of the reservoir is calculated using the CO2 flooding net present value model, and the injection and production plan, the calculated net present value, and the CO2 storage volume are stored in the database to provide basic data support for the subsequent proxy model construction.

[0073] Step 3: Divide the samples in the database into training set and test set, where samples form the training set, and the remaining The test set is composed of samples; a radial basis function proxy model is constructed based on the Gaussian kernel function and the inverse multi-quadratic kernel function, and a Gaussian process regression model is constructed based on the auto-correlation square exponential kernel function, and the samples in the training set are combined to construct a proxy model for each optimization target.

[0074] Step 4: Use the samples in the test set to obtain the coefficient of determination of the proxy models corresponding to the net present value of the reservoir and the carbon dioxide storage capacity. Based on the coefficient of determination, select the proxy model with the best prediction accuracy for each optimization target in the current optimization stage. That is, select the proxy model with a coefficient of determination close to 1 for each optimization target. Determine the adaptive proxy model for each optimization target in the current optimization stage and retrain it based on the database.

[0075] Step 5: Generate offspring populations using genetic algorithms and evaluate offspring individuals based on the proxy models of each optimization objective.

[0076] In this embodiment, when generating offspring population based on genetic algorithm, the offspring offset degree control parameter is set. 0.5, controls the crossover distribution index is 20, the mutation probability is 1, the mutation distribution index is 20, and the adaptive surrogate model of each optimization objective is used for evaluation.

[0077] Step 6: Construct a hypervolume-guided search strategy selection mechanism to select potential solutions based on the hypervolume evaluation index.

[0078] In this embodiment, first determine whether it is necessary to update the hypervolume improvement set in a stage If an update is required, the dominant search strategy, decomposition search strategy and index search strategy are traversed to select multiple potential solutions. Then, numerical simulations are performed on each selected potential solution, and the hypervolume improvement value is used to quantify the potential of each search strategy to advance the Pareto frontier. By calculating the hypervolume difference between the Pareto frontier and the reference point before and after optimization, the improvement of the search strategy on the convergence and diversity of the solution is intuitively evaluated, and the stage initialization of the hypervolume improvement set is completed; if the stage update of the hypervolume improvement set is not required, , then select the hypervolume improvement set The search strategy with the largest average value selects multiple potential solutions, performs numerical simulations based on each potential solution, and updates the database , calculate the current Pareto frontier relative to the previous optimized Pareto frontier super volume improvement value , update the hypervolume improvement set .

[0079] Step 7: Use the updated database Continue iterative optimization and repeat steps 3 to 6 to perform numerical simulation on the optimal potential solution until the number of reservoir numerical simulations reaches the preset total number of evaluations MaxFEs. Each iterative optimization records the evaluation samples and the corresponding reservoir net present value and carbon dioxide storage capacity to obtain an updated database. .

[0080] Step 8: Get the database The net present value and carbon dioxide storage capacity of the reservoir are obtained by numerical simulation evaluation of each sample. After the iterative optimization is completed, the reservoir numerical simulator outputs the sample and the target value corresponding to each iterative optimization, and obtains the non-dominated solution set based on the dominance relationship of the solution , obtain the Pareto frontier solution and its corresponding target results, and complete the multi-objective optimization of carbon dioxide flooding and storage.

[0081] To validate the beneficial effects of the present method, the present method, ASHMO, and the traditional multi-objective optimization method, KRVEA, were applied to the study area of this example. A comparison between the traditional Kriging-based multi-objective optimization method, KRVEA, and the present method, ASHMO, further verified the accuracy of the present method's adaptive surrogate model framework. The ASRVEA method, a reference vector-guided evolutionary algorithm based on an adaptive surrogate model, utilizes population distribution and angle-penalized distance to perform population search. Comparing the test results of the ASRVEA and ASHMO methods further validated the effectiveness of the hypervolume-guided search mechanism in the present method. The NSGA-III method, a hybrid multi-objective optimization algorithm, ranks populations based on dominance relationships and uses reference points to guide the search process. Comparing the test results of the NSGA-III and ASHMO methods demonstrates the superiority of the present method. The KTA2 method, a recently proposed multi-objective optimization algorithm for expensive optimization problems, balances population convergence and diversity by constructing a dual-archive structure. Comparing the test results of the KTA2 and ASHMO methods further validated the effectiveness of the present method for solving expensive optimization problems.

[0082] The above algorithms are set with the same parameters: the initial sampling sample size is set to 100, the initial population size is set to 50, the maximum number of numerical simulations is set to 1000, and gene exchange and mutation are performed using simulated binary crossover and polynomial mutation. The crossover probability and mutation probability are both set to 1, and the crossover distribution index and mutation distribution index are both set to 20. The adaptive proxy model includes a radial basis function with a Gaussian kernel and an inverse multi-quadratic kernel, and a Gaussian process regression model with an autocorrelation squared exponential kernel function. In addition, in the ASHMO method of the present invention, the number of hypervolume phase updates is set to 20, the number of proxy phase updates is set to 20, the angle penalty change rate exponent is set to 2, the evaluation sample size is set to 5, and the number of built-in evolutionary iterations of each search method is set to 1.

[0083] After the optimization is completed, the Pareto frontier of each method maximizing the carbon dioxide storage capacity and the net present value of the reservoir is as follows: Figure 5 As shown, from Figure 5 It can be seen that under the same number of numerical simulations, the Pareto front generated by the ASHMO method of the present invention has the best convergence and diversity in the target space. The performance of the proxy model of the ASHMO method and the KRVEA method in the late optimization stage is as follows: Figure 6 As shown. Figure 6 It can be seen that the fitting accuracy of the adaptive proxy model adopted by the method of the present invention is better.

[0084] Figure 7 The search strategy calls in the optimization process of this embodiment are as follows: Figure 7 It can be seen that the index search method is used the most times and the dominant search method is used the least times, indicating that the ASHMO method of the present invention can fully utilize the advantages of each search strategy in the entire optimization process. Further, the method of the present invention is applied to four typical strategy schemes, among which the first typical strategy scheme aims to maximize the carbon dioxide storage capacity, the second typical strategy scheme focuses on the carbon dioxide storage capacity, the third typical strategy scheme focuses on the net present value of the reservoir, and the fourth typical strategy scheme aims to maximize the net present value of the reservoir. The production conditions of the four typical strategy schemes are as follows: Figure 8 As shown, from Figure 8 As can be seen from the figure, the cumulative gas injection and water production of the four production schemes are basically the same, which shows that when the gas injection volume is basically the same, the CO2 recovery efficiency or storage efficiency can be improved by adjusting the injection-production system, which further reveals the differential regulation mechanism of the injection-production strategy on the development target in the multi-objective collaborative optimization. The carbon dioxide volume molar concentration and remaining oil field diagram of the four typical strategy schemes are shown in the figure. Figure 9 and Figure 10 As shown in the figure, the first typical strategy resulted in the highest mean molar CO2 concentration in the reservoir, but its uniformity was poor, with higher concentrations near the two injection wells and areas of lower concentration within the reservoir, corresponding to a higher amount of residual oil. In contrast, the fourth typical strategy resulted in the lowest residual oil saturation and molar CO2 concentration, but its distribution was more balanced, resulting in a more stable storage effect.

[0085] comprehensive Figures 5 to 10 It can be found that the multi-objective optimization method for carbon dioxide flooding and storage based on the adaptive agent model adopted in this embodiment can effectively balance the carbon dioxide flooding and storage injection and production optimization problem by weighing the optimization effects of the two objectives. Decision makers can obtain a set of Pareto solutions without distinguishing between good and bad, effectively avoiding the excessive pursuit of a certain goal that causes large losses to other goals.

[0086] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A multi-objective optimization method for CO2 flooding and storage based on an adaptive agent model, characterized by: The following steps are involved: Step 1: Based on the oilfield geological information and oilfield production information, a CO2 flooding component model is constructed and imported into a reservoir numerical simulator; Step 2: Build a CO2 flooding storage and injection-production optimization model. Use Latin hypercube sampling to obtain multiple injection-production scenarios to construct an initial sample set. Use a reservoir numerical simulator to calculate the net present value and CO2 storage capacity corresponding to each sample in the initial sample set, and then build a database. Step 3: Divide the samples in the database into training sets and test sets, and build proxy models for each optimization objective based on the samples in the training set; Step 4: Use the samples in the test set to obtain the root mean square error of the proxy model corresponding to each optimization target. Based on the determination coefficient, select the proxy model with the best prediction accuracy for each optimization target in the current optimization stage, and adapt the proxy model of each optimization target in the current optimization stage. Step 5: Generate offspring populations using genetic algorithms and evaluate offspring individuals based on the proxy models of each optimization objective; Step 6: Construct a hypervolume-guided search strategy selection mechanism to select potential solutions based on the hypervolume evaluation index. Step 7: Repeat steps 3 to 6 to perform numerical simulation on the optimal potential solution until the preset total number of evaluations MaxFEs is reached, and the updated database is obtained. ; Step 8: After the iterative optimization is completed, the reservoir numerical simulator outputs the samples of each iterative optimization and the corresponding reservoir net present value and CO2 storage capacity, obtains the Pareto front solution and the corresponding target results, and completes the multi-objective optimization of CO2 flooding and storage.

2. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 1, characterized in that: In step 1, a three-dimensional geological model is constructed based on oilfield geological data using oilfield geological modeling software, and the three-dimensional geological model is imported into a reservoir numerical simulator. Then, based on oilfield production information, the numerical simulation software is used to simulate and obtain high-pressure fluid physical properties and oil-gas-water relative permeability curves, and a fluid model is determined. The three-dimensional geological model and the fluid model are combined to construct a carbon dioxide flooding component model, and the carbon dioxide flooding component model is imported into the reservoir numerical simulator.

3. The multi-objective optimization method for carbon dioxide flooding and storage based on the adaptive agent model according to claim 2 is characterized in that: The step 2 includes the following sub-steps: Step 2.1: For the optimization problem of CO2 flooding and storage and injection-production, a CO2 flooding net present value model is constructed with maximizing the net present value of the reservoir and maximizing the CO2 storage volume as the dual objective function, and the development and control parameters of production wells and gas injection wells as the decision variables, combined with the variable boundary constraints. The net present value model of carbon dioxide flooding is: ； The constraints are: ； Where, is the economic net present value; is the time step number; is the reservoir production time; For the time steps; is the asset discount rate; is the well number; is the number of producing wells; For the Time step The oil production rate of the production wells; is the crude oil price; For the Time step Gas production rate of production wells; To process the produced gas cost; For the Time step The water production rate of the production well; is the produced fluid treatment cost; is the number of injection wells; For the Time step Gas injection rate of the injection well; is the gas injection cost; is the bottom hole flowing pressure of the production well; It is the lower limit of bottom hole flowing pressure of production well; The upper limit of the bottom hole flowing pressure of the production well; is the gas injection rate of the injection well; is the lower limit of the gas injection rate of the injection well; is the upper limit of the gas injection rate of the injection well; Step 2.2, based on the change in carbon dioxide concentration within the grid of the carbon dioxide flooding component model in the reservoir numerical simulator before and after injection and production, calculate the total carbon dioxide storage capacity and construct a carbon dioxide flooding storage capacity model; The carbon dioxide landfill inventory model is: ； Where, is the amount of carbon dioxide stored; is the grid number; is the total number of grids; For the The volume of the grid; Before injection The molar concentration of carbon dioxide per grid; After injection and production The molar concentration of carbon dioxide per grid; is the molar mass of carbon dioxide; Step 2.3: Combine the CO2 flooding net present value model and the CO2 flooding storage model to establish a CO2 flooding oil storage injection and production optimization model and import it into the reservoir numerical simulator; Step 2.4: Use Latin hypercube sampling to obtain multiple samples within the decision variable constraint domain and construct the initial sample set. , ,in, For the The injection plan of each sample, is the initial sample size; Step 2.5: Import the initial sample set into the reservoir numerical simulator, and use the reservoir numerical simulator to calculate the net present value of the reservoir and the carbon dioxide storage capacity of each sample, and obtain the initial sample set. Corresponding dual-target dataset ,in, is the reservoir net present value dataset, , For the The net present value of the reservoir of the sample, is the carbon dioxide storage data set, , For the The amount of CO2 stored in each sample; Step 2.6, based on the initial sample set and its corresponding reservoir net present value dataset and CO2 storage datasets , build the database , .

4. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 3, characterized in that: Step 3 includes the following sub-steps: Step 3.1: Get the total sample size of the database , randomly selected The samples are used to form a training set, and the remaining samples are used to form a test set; the training set includes a reservoir net present value training set and CO2 storage training set , where the reservoir net present value training set , CO2 storage training set ; Step 3.2: constructing a proxy model based on the samples in the training set, wherein the proxy model includes a radial basis function proxy model and a Gaussian process regression model. The radial basis function proxy model is constructed based on a Gaussian kernel function and an inverse multi-quadratic kernel function, and the Gaussian process regression model is constructed based on an autocorrelation squared exponential kernel function. The radial basis function proxy model is: ； in, When the kernel function is a Gaussian kernel function, the expression of the kernel function is: ； When the kernel function is an inverse multi-quadratic kernel function, the expression of the kernel function is: ； Where, is the radial basis function; is the Euclidean distance between the input point and the center point; is the serial number of the kernel function; is the total number of kernel functions; For the The weight of the kernel function; is the kernel function; is the bias term; is the shape parameter used to determine the function expansion; is the shape parameter used to determine the smoothness of the function; The Gaussian process regression model is: ； in, The autocorrelation squared exponential kernel function for: ； Where, 、 All are input points; is the fitting function; is the Gaussian process function; is the mean of the distribution; is the kernel function used to measure the similarity between any input points; To calculate the expected value of a random variable; is the signal variance; is an exponential function; is the dimension number; is the input dimension; is the length scale of the input dimension; For input point No. The value of the dimensional input variable; For input point No. The value of the input variable.

5. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 4, characterized in that: The step 4 includes the following sub-steps: Step 4.1: Substitute the reservoir net present value and CO2 storage capacity of each test sample in the test set into a radial basis function proxy model with a Gaussian kernel, an inverse multi-quadratic kernel function, and a Gaussian process regression model with an autocorrelation squared exponential kernel function, respectively, and determine the coefficient of determination corresponding to each proxy model; The determination coefficient calculation formula is: ； Where, is the coefficient of determination; is the test sample serial number; is the true value of the surrogate model; is the predicted value of the surrogate model; is the average value of the true value of the test sample; In step 4.2, the proxy model with the determination coefficient closest to 1 among the proxy models is selected as the best proxy model for the optimization target, and the proxy model with the best prediction accuracy for each optimization target in the current optimization stage is obtained, and the proxy model of each optimization target in the current optimization stage is adapted to suit the needs of the optimization target.

6. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 5, characterized in that: The step 5 includes the following sub-steps: Step 5.1, set the number of updates and potential samples of the adaptive agent model stage, set the total number of reservoir numerical simulation evaluations MaxFEs, and use the genetic algorithm to generate the offspring population; The process of generating offspring population by the genetic algorithm includes gene exchange and gene mutation; The gene exchange process is: ； ； The gene mutation process is: ； Where, is the individual serial number, is the gene number, is the total number of genes; The first individual; The parent individual 1 genes; The parent individual 2 genes; Control parameters for the degree of offspring offset; is a random number; is the distribution index for controlling crossover; The first a new individual; The first individual; is the disturbance factor, which follows a multinomial distribution; is the upper limit of the variable; is the lower limit of the variable; In step 5.2, the proxy model of each optimization objective is used to predict the net present value of the oil reservoir and the carbon dioxide storage capacity corresponding to the individuals in the offspring population.

7. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 6, characterized in that: The step 6 includes the following sub-steps: Step 6.1, set the number of hypervolume stage updates and determine whether the hypervolume improvement set needs to be updated in stages If an update is required, the dominance search strategy, decomposition search strategy and index search strategy are traversed to optimize the potential solution; The dominance search strategy uses a fast non-dominated sorting genetic algorithm to calculate the population dominance relationship, selectively sorts individuals in the same non-dominated layer based on their crowding distance, and selects individuals with larger crowding distance as potential solutions; The decomposition search strategy decomposes the high-dimensional multi-objective optimization problem into several single-objective sub-problems, enhances the control ability of the objective trade-off relationship through reference points or weight vectors, and guides the evolutionary algorithm to select potential solutions based on the reference vectors and update the reference vectors; The index-based search strategy uses performance indicators to quantitatively evaluate individuals in the population, and balances population diversity and algorithm convergence through fitness value update and environment selection mechanism. By calculating the updated fitness values of all individuals in the population and sorting them, the current worst individual is deleted and the population is updated, and the fitness value of the updated population is calculated until a potential solution is screened out; Step 6.2, perform numerical simulation on each potential solution selected, and initialize the hypervolume improvement set , ,in, The hypervolume increment of the dominant search strategy, For decomposition class search strategy hypervolume increment, Search strategy for indicator class super volume increment; The hypervolume improvement value is used to quantify the potential of each search strategy to advance the Pareto frontier. By calculating the hypervolume difference between the Pareto frontier and the reference point before and after optimization, the improvement of the search strategy on solution convergence and diversity is intuitively evaluated. The hypervolume improvement set is initialized. ,get: ； Where, For the The improvement value of the hypervolume after iterations; For the The set of non-dominated solutions after iterations; is the hypervolume calculation function; is the number of iterations; For the The reference points generated after iterations; Step 6.3, if no stage is required to update the hypervolume improvement set , then select the hypervolume improvement set The search strategy with the largest average value selects the potential solution and performs numerical simulation to update the database. , calculate the current Pareto frontier relative to the previous optimized Pareto frontier super volume improvement value , update the hypervolume improvement set, the updated hypervolume improvement set is: ； in, For the The hypervolume increment of the dominant search strategy after iterations, For the The decomposition search strategy hypervolume increment after iterations, For the The hypervolume increment of the indicator-based search strategy after iterations.

8. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 7, characterized in that: The individual crowding calculation formula in the dominant search strategy is: ； Where, For individuals crowding distance; is the ordinal number of the target dimension; is the total number of target dimensions; For individuals In the target dimension Target value on Target dimension Upper maximum value; Target dimension Upper minimum value; The process of optimizing potential solutions by the decomposition search strategy is as follows: First, the target value is normalized based on the ideal point to obtain: ； Secondly, the population is divided and the individuals are assigned to the subpopulation corresponding to the expected closest reference vector, and the result is: ； Finally, based on the convergence and diversity of the angle-penalized distance balance solution, we obtain: ； Where, For individuals Target value before conversion; For individuals The target value after conversion; For the current population The minimum target value in dimensions; For the current population Target value for each dimension; is the reference vector the corresponding subpopulation; For the individual; For individuals The serial number of the reference vector corresponding to the maximum cosine angle between it and the reference vector; is the complex argument calculation function; is the maximum value function; is the cosine function; is the arccosine function; is the two-norm; is the serial number of the reference vector; is the total number of reference vectors; For individuals With reference vector The cosine angle between For the reference vectors; For individuals With reference vector The cosine angle between For individuals To the reference vector Angle penalty distance; is the penalty function; is the current iteration number; is the maximum number of iterations; is the rate of change control parameter of the penalty function; is the reference vector Minimum angle with other reference vectors; The calculation formula for updating the reference vector in the decomposition search strategy is: ； Where, For the After the adjustment reference vectors; The first reference vectors; is the element-wise product of vectors; For the The maximum value of each objective function in the generation; For the The minimum value of each objective function in the generation; The quantitative evaluation process of the individual population by the indicator search strategy is as follows: ； ； ； Where, is the individual serial number; Calculate the function for the indicator value between two individuals; is the normalized individual The value of the optimization target; To optimize the target sequence number; is the total number of optimization targets; For individuals in The original value of the optimization target; is the minimum function; is the individual fitness value; is the total number of individuals in the population; is the normalized index value; is the scaling factor; is the fitness value of the individual after update.

9. The multi-objective optimization method for carbon dioxide flooding and storage based on an adaptive agent model according to claim 8, characterized in that: In step 8, after the iterative optimization is completed, the non-dominated solution set is obtained based on the dominance relationship of the solution. : ； Where, 、 All are Chinese databases individuals in It is a Pareto dominance relationship; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; For the database Medium individuals No. dimension; The Pareto solution set is the set of all Pareto optimal solutions in the decision space. The Pareto solution set constructed using non-dominated solutions is: ； Where, is the target value of all dimensions corresponding to the individual; For decision-making space.

Images