Lithologic reservoir variable credibility proxy optimization method based on multi-source multi-precision data fusion

Abstract

The present application relates to the technical field of lithologic reservoir scheme optimization, and discloses a multi-source multi-precision data fusion lithologic reservoir variable credibility agent optimization method, comprising: constructing a low-precision carbonate reservoir numerical model considering only pores and a high-precision carbonate reservoir numerical model considering both pores and fractures; setting optimization variables, designing a space, and designing a production scheme multi-objective optimization model with economic parameters and cumulative oil production as objective functions; based on the design space, using a Latin hypercube sampling method to sample the optimization variables to generate n l first production schemes and n h second production schemes; based on a scaling function, establishing a variable credibility agent model to fuse the objective functions calculated from the first production schemes and the second production schemes; and based on a hybrid particle swarm algorithm, optimizing each parameter in the production scheme multi-objective optimization model. The present application greatly improves optimization efficiency and the quality of optimization schemes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithologic reservoir scheme optimization, and particularly relates to a lithologic reservoir variable credibility agent optimization method based on multi-source multi-precision data fusion. BACKGROUND

[0002] In the commercial development of carbonate reservoirs, the optimization design of the development scheme directly determines the development benefit and the ultimate recovery rate. Due to the complex pore structure and permeability distribution of carbonate reservoirs, the development of carbonate reservoirs is more difficult, and the production behavior of the reservoir is difficult to predict. Therefore, it is crucial to develop an efficient development scheme. The traditional optimization design of the development scheme of carbonate reservoirs often relies on the experience of reservoir engineers, which can provide a preliminary direction and strategy for development to some extent. However, relying solely on experience often cannot fully grasp the complex physical property changes of the reservoir and their impact on oil production.

[0003] In order to improve the accuracy and reliability of the design, reservoir numerical simulation technology has been widely applied in the design of reservoir development schemes. It can predict the production behavior and recovery rate of the reservoir under different development schemes by establishing a geological and reservoir model and using computer simulation. These models require a large amount of input data, such as rock physical parameters, fluid properties, well pattern arrangement, etc., and the simulation process requires a large amount of computing resources and time. Although numerical simulation can provide more accurate prediction results, the high computing cost and long computing period to some extent limit the breadth and efficiency of its application.

[0004] In recent years, with the development of artificial intelligence technology, reservoir agent model technology has been widely applied in the oil and gas industry and has become a key tool in reservoir engineering. The reservoir agent model learns and captures the complex behavior and mutual relationship of the reservoir from a large number of numerical simulation data through machine learning algorithms, so as to realize the rapid prediction and optimization of the reservoir development scheme. This technology can greatly reduce the computing cost, improve the speed and flexibility of the optimization design, so that engineers can evaluate the potential effect of multiple development schemes in a short time and conduct sensitivity analysis and uncertainty evaluation. With the introduction of advanced artificial intelligence technologies such as deep learning, the accuracy and adaptability of the agent model are continuously improved, which promotes the transformation of reservoir development from experience-driven to data-driven and intelligent decision-making, and opens up new prospects for improving the efficiency and recovery rate of oilfield development. However, carbonate reservoirs have complex reservoir pore structure, strong heterogeneity of physical properties, and multi-scale flow mechanisms, which make it difficult for existing agent modeling methods to accurately capture their flow characteristics. Therefore, it is urgent to develop new agent optimization methods and frameworks for carbonate reservoir development to realize efficient optimization design of carbonate reservoir production schemes. SUMMARY

[0005] The application provides a lithologic reservoir variable credibility agent optimization method based on multi-source and multi-precision data fusion, and aims to solve the problem that a huge amount of calculation resources are required for traditional carbonate reservoir production optimization and it is difficult to find a global optimal solution.

[0006] The application is achieved by the following technical solutions:

[0007] A lithologic reservoir variable credibility agent optimization method based on multi-source and multi-precision data fusion comprises the following steps:

[0008] A low-precision carbonate reservoir numerical model considering only pores and a high-precision carbonate reservoir numerical model considering both pores and fractures are constructed;

[0009] An optimization variable, a design space, and a production scheme multi-objective optimization model with economic parameters and cumulative oil production as target functions are set;

[0010] The optimization variable is sampled based on the design space by using a Latin hypercube sampling method to generate first production schemes and second production schemes, wherein the first production schemes are used to calculate the target functions of the low-precision carbonate reservoir numerical model, and the second production schemes are used to calculate the target functions of the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model;

[0011] A variable credibility agent model is established based on a scaling function to fuse the target functions calculated by the first production schemes and the second production schemes respectively;

[0012] Each parameter in the production scheme multi-objective optimization model is optimized based on a hybrid particle swarm algorithm, wherein in the optimization process, the variable credibility agent model is updated based on the particle individuals and the population in the hybrid particle swarm algorithm.

[0013] As optimization, the vector form of the optimization variable is represented as wherein the optimization variable is and , represents the injection rate of the injection well j, represents the bottom hole pressure of the production well p, and the design space comprises the injection rate range of the injection well and the bottom hole pressure range of the production well, wherein the injection rate range of the injection well is represented as , and the bottom hole pressure range of the production well is represented as wherein , respectively represent the lower limit value and the upper limit value of the injection rate of the injection well, , These represent the lower and upper limits of the bottom pressure level of the production well, respectively.

[0014] As an optimization, the objective function is:

[0015]

[0016] The constraints are:

[0017]

[0018] in, Indicates economic parameters, This indicates the cumulative oil production. g ( u ) is a constraint condition. lb and ub To optimize the lower and upper limits of variables, u To optimize the vector form of the variables;

[0019] , ;

[0020] in, For oil prices; The price is for water treatment. Gas processing price; This is the price for water injection; For gas injection price, 、 and For the first j Koujing in the n Oil production, water production, and gas production over a specific time period; and For the first j Koujing in the n Water and gas injection volumes for each time period; This refers to the number of production time steps. b The discount rate; P The number of production wells; I The number of injection wells; For production time steps, For the first n The length of each time step.

[0021] As an optimization, the optimization variables are sampled using the Latin hypercube sampling method based on the design space to generate... One and a first production scheme and a second production scheme, and calculating the objective function by the low-precision carbonate reservoir numerical model, and calculating the objective function by the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model, and the specific process is as follows:

[0022] A1. Generating a first production scheme:

[0023] A1.1. dividing the injection rate and the bottom hole pressure into intervals respectively based on the corresponding design space, so that the design space is divided into intervals;

[0024] A1.2. performing times of random sampling in the intervals, and each of the first production schemes formed by sampling includes an injection rate and a bottom hole pressure, so that finally first production schemes are obtained , denotes the i-th first production scheme, ;

[0025] A2. Generating a second production scheme is:

[0026] A2.1. dividing the injection rate and the bottom hole pressure into intervals respectively based on the corresponding design space, so that the design space is divided into intervals;

[0027] A2.2. performing times of random sampling in the intervals, and each of the second production schemes formed by sampling includes an injection rate and a bottom hole pressure, so that finally second production schemes are obtained , denotes the i-th second production scheme, ;

[0028] A3. calculating the objective function by the low-precision carbonate reservoir numerical model on the first production schemes, so that first objective function response values are obtained, and the i-th first objective function response value is expressed as ;

[0029] A4. calculating the objective function by the high-precision carbonate reservoir numerical model on the second production schemes, so that a second objective function response value, a second objective function response value is expressed as ;

[0030] A5, by the low-precision carbonate reservoir numerical model to a second production scheme objective function calculation to obtain a third objective function response value, a third objective function response value is expressed as .

[0031] As optimization, the variable credibility proxy model is expressed as:

[0032] ;

[0033] wherein, indicates a variable credibility proxy model, wherein a set of sampling data x including an injection volume and a bottom hole pressure, indicates a scale function characterized by a Gaussian process regression model , , , .

[0034] As optimization, The specific expression is:

[0035] ;

[0036] ;

[0037] is a correlation vector is a correlation matrix, is a column vector filled with 1 with a length of n h .

[0038] As optimization, the specific process of optimizing each parameter in the production scheme multi-objective optimization model based on the hybrid particle swarm algorithm is:

[0039] B1, initialize parameters, set the maximum number of iterations, the maximum speed of particle individuals, population size N and termination condition, and initialize the position and speed of the particle individuals to form a population, wherein the population is composed of several optimization variables;

[0040] B2, calculate the fitness value of each particle individual;

[0041] B3, according to the fitness value, select the global optimal value and the individual optimal value from the population ;

[0042] B4. Randomly select half of the particles in the population and perform mutation operations using the DE / rand-to-best operator. Simultaneously, randomly select particles in the population and perform crossover operations according to a set crossover operator. The crossover operator is specifically:

[0043]

[0044] in, Indicates the first The position of the i-th particle in the j-th dimension during the iteration. , Indicates the injection volume. Indicates the bottom hole pressure. Indicates the first The position of the j-th dimension of the i-th particle after mutation using the DE / rand-to-best operator in the iteration count. ,in, Scaling factor and It is a random number. N is a positive integer. Indicates the first The position of the i-th particle in the j-th dimension during the iteration count. The crossover probability;

[0045] Meanwhile, the other half of the particles in the population are updated according to the following formula:

[0046] ;

[0047] ;

[0048] ;

[0049] B5. Calculate the fitness value of each individual particle in the new population, and update the population using the selection operation operator, wherein the selection operation operator is specifically:

[0050] ;

[0051] in, Indicates the first The individual particles after iteration, Indicates the first The individual particles after iteration, Indicates the first The fitness value of an individual particle after iteration. Indicates the first fitness value of the particle individual, wherein the fitness value of the particle individual is an objective function of the particle individual;

[0052] B6, determining whether a termination condition is met, if yes, outputting a global optimal value , otherwise, jumping to B7;

[0053] B7, determining whether a predefined fixed iteration number is reached, wherein the fixed iteration number is less than a maximum iteration number, if yes, jumping to B8, otherwise, returning to B4;

[0054] B8, taking a minimum Euclidean distance between the particle individuals in the current population and samples in a high-precision sample set composed of a second production scheme other than the population as a new particle individual to update the population, and then jumping to B6.

[0055] As optimization, the fitness value of the particle individual is an objective function of the particle individual, and when calculating the fitness value of the particle individual, if the dominance status of the particle individual does not change due to the prediction uncertainty of the variable confidence proxy model, the fitness value of the particle individual is calculated by the variable confidence proxy model, otherwise, the fitness value is calculated by the high-precision carbonate reservoir numerical model.

[0056] As optimization, the representation of the prediction uncertainty of the particle individual is as follows:

[0057]

[0058] wherein, represents the i-th particle individual, if , it is determined that the dominance status of the particle individual does not change due to the prediction uncertainty of the variable confidence proxy model, wherein, ; , represents a dominated point, represents a non-dominated point, and , M is the number of particle individuals in the population, is the minimum value of the minimum distance between all dominated points and all non-dominated points, is a variance; is a correlation vector of a scaling function; is a correlation matrix of a scaling function, represents the prediction uncertainty of the particle individual.

[0059] As optimization, in B8, the corresponding formula for population updating is:

[0060]

[0061] wherein, is a high-precision sample set, l is the number of samples in the high-precision sample set, represents the Euclidean distance between the particle individual i of the current population and the mth sample in the high-precision sample set.

[0062] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0063] The present application is directed to the reservoir properties and pore deconstruction characteristics of carbonate reservoirs, and different precision numerical models of carbonate reservoirs are established by considering only pores and pores and fractures simultaneously, and on this basis, multi-source data is sampled to obtain data samples of different precisions, and a variable confidence proxy modeling method based on a scaling function is applied to realize the fusion proxy modeling of multi-source multi-precision data, to play the characteristics and advantages of different precision data samples and improve the prediction accuracy of the proxy model. Based on multi-source multi-precision data, variable confidence modeling technology based on a scaling function, and model iteration update strategy, the model update strategy based on individuals and populations is considered in the development scheme optimization iteration process, to realize efficient and accurate construction of the proxy model of the carbonate reservoir. On this basis, combined with a hybrid multi-objective particle swarm optimization algorithm, an intelligent optimization framework based on the variable confidence proxy model and the multi-objective optimization algorithm is built to realize efficient multi-objective optimization of the production scheme of the carbonate reservoir. The optimization method and framework can complete the prediction of the production scheme of the carbonate reservoir within seconds, realize the rapid response of the objective function in the production scheme optimization iteration process, replace time-consuming numerical simulation calculation, and overcome the problem that the traditional proxy model cannot adapt to the complex reservoir characteristics of the carbonate reservoir, greatly improving the optimization efficiency and the quality of the optimization scheme. BRIEF DESCRIPTION OF DRAWINGS

[0064] The drawings described herein are used to provide further understanding of the embodiments of the present application, constitute a part of the present application, and do not constitute a limitation on the embodiments of the present application. In the drawings:

[0065] Figure 1 is a flowchart of the optimization method of the present application;

[0066] Figure 2 is a low-precision numerical simulation model interface diagram;

[0067] Figure 3 is a high-precision numerical simulation model interface diagram;

[0068] Figure 4 is a variable confidence model NPV prediction accuracy verification diagram;

[0069] Figure 5 is a variable confidence model cumulative oil production prediction accuracy verification diagram;

[0070] Figure 6 A schematic diagram of a multi-objective optimization solution set in the embodiment;

[0071] Figure 7 A display diagram of a production scheme with relatively large NPV and cumulative oil production in the embodiment;

[0072] Figure 8 A schematic diagram of the calculation process of MMD in the embodiment;

[0073] Figure 9 A representation diagram of the relationship between two MMDs and prediction bias in the embodiment. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be given below in combination with embodiments and drawings. The schematic embodiments of the present application and the description thereof are only used to explain the present application, and do not limit the present application.

[0075] The embodiment 1 provides a lithologic reservoir variable credibility agent optimization method of multi-source multi-precision data fusion, comprising:

[0076] S1, constructing a low-precision carbonate reservoir numerical model considering only pores and a high-precision carbonate reservoir numerical model considering pores and fractures at the same time;

[0077] That is, this step is to build a sampling numerical model. The different precision carbonate rock reservoir numerical models A and B considering only pores and considering pores and fractures at the same time are constructed, and the grid number n , the fracture number n f and the reservoir scale x × y × z are set according to the actual situation of the reservoir.

[0078] That is, in the embodiment, the model A is a low-precision carbonate reservoir numerical model, and the model B is a high-precision carbonate reservoir numerical model.

[0079] S2, setting optimization variables, design space, and a production scheme multi-objective optimization model with economic parameters and cumulative oil production as objective functions;

[0080] That is, this step is to design optimization variables, design space, objective functions and a production scheme multi-objective optimization model.

[0081] 1, optimization variable definition, setting the injection rate of the injection well I j , the bottom hole pressure of the production well P p, the optimization variable vector u =( I j , P p )。

[0082] i.e. the vector form of the optimization variable is where the optimization variable is and , represents the injection rate of the injection well j, represents the bottom hole pressure of the production well p, j is the number of the injection well, p is the number of the production well.

[0083] 2. Define the design space of the optimization variable, set the injection rate range of the injection well as I min to I max m 3 / day, and the bottom hole pressure range of the production well as P min to P max MPa.

[0084] i.e. the design space includes the injection rate range of the injection well and the bottom hole pressure range of the production well, wherein the injection rate range of the injection well is represented as and the bottom hole pressure range of the production well is represented as wherein , respectively represent the lower limit value and the upper limit value of the injection rate of the injection well, , respectively represent the lower limit value and the upper limit value of the bottom hole pressure of the production well.

[0085] Define the objective function, consider the economic parameters NPV ( f 1) and cumulative oil production COP ( f 2) as the objective function of multi-objective optimization:

[0086] (1),

[0087] (2);

[0088] wherein is the oil price; is the water treatment price; is the gas treatment price; is the injection water price; is the injection gas price, 、 and is the number of the first time period j is the oil production, water production and gas production of the first well in the first time period; n is the water injection and gas injection of the first well in the first time period; and is the number of the first time period j is the water injection and gas injection of the first well in the first time period; n is the number of the production time step; is the discount rate; b is the number of the production well; P is the number of the injection well; I is the production time step, is the length of the first time step. n 4. The production scheme multi-objective optimization model is based on the optimization variable, the design space and the target function setting, and establishes the production scheme multi-objective optimization model as follows:

[0089] 4. The production scheme multi-objective optimization model is based on the optimization variable, the design space and the target function setting, and establishes the production scheme multi-objective optimization model as follows:

[0090] (3)

[0091] The constraint condition is:

[0092] (4)

[0093] wherein, indicates the economic parameter, indicates the cumulative oil production, g ( u ) is the constraint condition, lb and ub are the lower limit value and the upper limit value of the optimization variable, u is the vector form of the optimization variable; , .

[0094] S3, based on the design space, a Latin hypercube sampling method is used to sample the optimization variable to generate first and second production schemes, wherein the first production scheme is used for the low-precision carbonate reservoir numerical model to calculate the target function, and the second production scheme is used for the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model to calculate the target function; For example, the production scheme is equivalent to x, the target function is equivalent to y, and the carbonate reservoir numerical model is equivalent to the function of the relationship between x and y. The production method is input into the carbonate reservoir numerical model, simulation calculation is performed, and the target function value is obtained.

[0095]

[0096] ​​The Latin hypercube sampling method (LHS) is used for multi-source, multi-precision data sampling. Within the design space of the optimization variables, the LHS method is applied to generate... n l and n h The basic principle of this production plan is as follows: Assume that in a... k Generation of dimensional design variable space n One data sample, that is, to perform n The second sampling requires... k Each design variable is equally divided within its defined range. n This divides the entire design space into smaller intervals, thus creating an even division. n ) k Each interval is defined, and then within the design space, intervals from 0, 1, ... are defined. n -1 n A sampling method that meets the following conditions is valid: (1) the sample points in each sampling are randomly distributed within equally divided intervals; (2) the projections of all data sample points on any dimension are unique after sampling. The A model is then used to analyze... n l The objective function of each production plan is calculated, and A and B models are used respectively. n h The objective function is calculated for each production plan.

[0097] In some embodiments, the optimization variables are sampled using the Latin hypercube sampling method based on the design space to generate... One and The specific process of calculating the objective function using the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model, based on the first and second production schemes, is as follows:

[0098] A1, Generation The first production plan:

[0099] A1.1 Divide the injection volume and bottom hole pressure into equal portions based on the corresponding design space. The design space is divided into several small intervals, thus making the design space evenly divided. Small intervals;

[0100] A1.2, in In each small interval Each random sampling, and the first production plan formed in each sampling, includes an injection volume and a bottom hole pressure, thus ultimately obtaining... The first production plan , This represents the i-th first production plan. ;

[0101] A2、generating a second production scheme is:

[0102] A2.1、dividing the injection rate and the bottom hole pressure into small intervals respectively based on the corresponding design space, so that the design space is divided into small intervals;

[0103] A2.2、in small intervals, performing times of random sampling, each time of sampling forming a second production scheme including an injection rate and a bottom hole pressure, so as to finally obtain second production schemes , the i-th second production scheme is denoted as ;

[0104] A3、calculating the objective function of first production schemes by the low-precision carbonate reservoir numerical model, so as to obtain first objective function response values, the i-th first objective function response value is denoted as ;

[0105] A4、calculating the objective function of second production schemes by the high-precision carbonate reservoir numerical model, so as to obtain second objective function response values, the i-th second objective function response value is denoted as ;

[0106] A5、calculating the objective function of second production schemes by the low-precision carbonate reservoir numerical model, so as to obtain third objective function response values, the i-th third objective function response value is denoted as .

[0107] S4、based on the scaling function, establishing a variable credibility proxy model for fusing the objective functions calculated by the first production schemes and the second production schemes respectively;

[0108] The variable credibility surrogate model is used to realize the acceleration of the target function calculation in the optimization process, instead of the numerical simulation process. The purpose of the variable credibility is to realize a dynamic balance between the modeling calculation burden and the model performance by fusing data of different precisions, and to improve the performance of the surrogate model under limited computing resources.

[0109] That is, the variable credibility modeling of multi-source multi-precision data fusion based on the scaling function. The scaling function is expressed as the gap between the high-precision model and the low-precision model, and the high-precision data sample set is taken as the benchmark, and the high-precision target function response is The sample set of the scaling function is The scaling function in the variable credibility surrogate model is expressed as follows:

[0110] (11)

[0111] The Gaussian process regression model is used to represent the scaling function as follows:

[0112] (12)

[0113] The detailed Gaussian process regression model can be referred to the related literature.

[0114] On this basis, the variable credibility surrogate model based on the scaling function is expressed as follows:

[0115] (13)

[0116] Based on the above process, in the embodiment, the variable credibility surrogate model is expressed as:

[0117] ;

[0118] Wherein, represents the variable credibility surrogate model, wherein a set of sampling data x includes an injection volume and a bottom hole pressure, represents the scaling function represented by the Gaussian process regression model , , , .

[0119] The specific expression of is as follows:

[0120] ;

[0121] ;

[0122] is a correlation vector is a correlation matrix, is a column vector filled with 1 with length n h .

[0123] S5, based on the hybrid particle swarm algorithm, the production scheme multi-objective optimization model in each parameter optimization, wherein, in the optimization process, based on the particle individual and population update the variable credibility proxy model in the hybrid particle swarm algorithm.

[0124] That is, this step is based on the hybrid particle swarm algorithm DE-PSO algorithm combined with variable credibility proxy model of multi-objective optimization.

[0125] The pseudo code of the conventional DE-PSO algorithm is as follows:

[0126] Initialize population P and velocity V

[0127] Calculate the fitness of population P

[0128] Set the global best position g_best and individual best position p_best

[0129] While the termination condition is not met:

[0130] For each individual i in population P:

[0131] If i belongs to the differential evolution part:

[0132] Generate new individuals u_i through DE mutation and crossover

[0133] Else:

[0134] Generate new positions x_i through PSO velocity and position update

[0135] Calculate the fitness of new individuals

[0136] Update p_best and g_best according to the fitness

[0137] Update the population using the selection operation

[0138] Output the global best solution g_best

[0139] DE-PSO can effectively balance the global search and local search ability by combining the advantages of DE and PSO.

[0140] In the present application, the update of particle individual and population sets a new strategy. Next, step S5 is specifically introduced.

[0141] 1. Initialization:

[0142] B1. Initialize parameters, set the maximum number of iterations, the maximum velocity of individual particles, the population size N, and the termination condition, and initialize the position and velocity of the individual particles to form a population, wherein the population consists of several optimization variables;

[0143] The first and second production schemes are used to construct the variable credibility proxy model. Therefore, the initialization of individual particles here is based on a randomly optimized production scheme, which is randomly generated and selected, unlike the first and second production schemes mentioned earlier. Since the sampling parameters are continuous variables, they are directly initialized using the following formula: ,in, and These are the lower and upper bounds for each dimension of the optimization variables, respectively; Indicates on the interval Random numbers.

[0144] B2. Calculate the fitness value of each individual particle;

[0145] B3. Select the globally optimal value from the population based on the fitness value. and individual optimal value ;

[0146] 2. Perform the mixing operation:

[0147] B4. Randomly select half of the particles in the population and perform mutation operations using the DE / rand-to-best operator. Simultaneously, randomly select particles in the population and perform crossover operations according to a set crossover operator. The crossover operator is specifically:

[0148]

[0149] in, Indicates the first The position of the i-th particle in the j-th dimension during the iteration count. , Indicates the injection volume. Indicates the bottom hole pressure. Indicates the first The position of the j-th dimension of the i-th particle after mutation using the DE / rand-to-best operator in the iteration count. ,in, Scaling factor and It is a random number. N is a positive integer. Indicates the first The position of the i-th particle in the j-th dimension during the iteration count. The crossover probability;

[0150] At the same time, the other half of the particle individuals in the population are updated according to the following formula:

[0151]

[0152]

[0153]

[0154] That is, the mixing operation mainly includes DE mutation and crossover operation and PSO speed and position update operation:

[0155] ① DE mutation and crossover operation:

[0156] The differential evolution algorithm (DE) realizes individual mutation through differential strategy. The common differential strategy is to randomly select two or more individuals in the population, calculate the vector difference, and further sum up after scaling the vector difference. According to the difference of mutation operator, there are many mutation modes of differential evolution algorithm. DE / rand algorithm has strong global search ability, but the convergence speed is slow. DE / best has strong local search ability and fast convergence speed, but it is easy to fall into local optimum.

[0157] Randomly select half of the individuals in the population, and use the DE / rand-to-best operator combined with DE / rand and DE / best to balance the global and local search ability of the algorithm: Wherein, and are random numbers, is called a scaling factor, which is a certain constant. The purpose of the crossover operation is to randomly select individuals, because differential evolution is also a random algorithm, and the crossover operation operator adopted by the differential evolution algorithm is: Wherein, is the crossover probability; represents the generation.

[0158] ② PSO speed and position update

[0159] For the other part of the individuals in the population, the speed and position update formula of PSO is used to generate new individuals. In a n dimensional space, there are m particles, and the position t of particle i at time t is . The speed of particle i is . At time t , the position of particle i in the​​​j The velocity and position updates in each dimension are as follows:

[0160]

[0161]

[0162] .

[0163] B5. Calculate the fitness value of each individual particle in the new population, and update the population using the selection operation operator, wherein the selection operation operator is specifically:

[0164] ;

[0165] in, Indicates the first The individual particles after iteration, Indicates the first The individual particles after iteration, Indicates the first The fitness value of an individual particle after iteration. Indicates the first The fitness value of the individual particle after iteration; wherein, the fitness value of the individual particle is the objective function of the individual particle; according to the greedy selection strategy in DE, that is, selecting the better individual as the new individual, for the maximization problem, the selection operator is: .

[0166] B6. Determine if the termination condition is met. If so, output the global optimum. Otherwise, navigate to B7;

[0167] B7. Determine whether the predefined fixed number of iterations has been reached, wherein the fixed number of iterations is less than the maximum number of iterations. If yes, jump to B8; otherwise, return to B4.

[0168] B8. The population is updated by taking the minimum Euclidean distance between the particle individuals in the current population and the samples in the high-precision sample set composed of the second production scheme other than the population as the new particle individuals, and then the process jumps to B6.

[0169] Specifically, the fitness value of the individual particle is the objective function of the individual particle. When calculating the fitness value of the individual particle, if the dominance state of the individual particle does not change due to the prediction uncertainty of the variable confidence surrogate model, then the fitness value of the individual particle is calculated by the variable confidence surrogate model; otherwise, it is calculated by the high-precision carbonate reservoir numerical model.

[0170] The prediction uncertainty of the individual particles is characterized as follows:

[0171]

[0172] is the variance; is the correlation vector of the scaling function; is the correlation matrix of the scaling function;

[0173] wherein, denotes the ith particle individual, if , and M is the number of particle individuals in the population, then the dominance status of the particle individual is determined not to have changed due to the prediction uncertainty of the variable confidence surrogate model, wherein, ; , denotes a dominated point, denotes a non-dominated point, is the minimum value of the minimum distance between all dominated points and all non-dominated points, denotes the prediction uncertainty of the particle individual.

[0174] It should be noted that the dominated point (Dominated point) and the non-dominated point (Non-dominated point) are basic concepts for evaluating the relative merits of the solution set.

[0175] Dominance: In multi-objective optimization, a solution is said to dominate another solution if is not worse than in all objectives, and strictly better than in at least one objective. That is, is equivalent to in all objectives, but better than in at least one objective.

[0176] Non-dominance: If no solution can dominate another solution, the solution is non-dominated. In simple terms, x1 is non-dominated, and there is no solution that can be better than it in all objectives.

[0177] That is, the dominated point is worse than the non-dominated point and is dominated by the non-dominated point.

[0178] (or called ), denotes a solution in multi-objective optimization, then is a solution that is not worse than (dominate ).

[0179] The population update corresponds to the formula:

[0180]

[0181] wherein, is a high-precision sample set, l is the number of samples in the high-precision sample set, represents the Euclidean distance between the particle individual i of the current population and the mth sample in the high-precision sample set.

[0182] Next, the specific process of updating the variable confidence proxy model based on the particle individual and population in the hybrid particle swarm algorithm will be introduced.

[0183] The update strategy of the variable confidence proxy model based on individuals;

[0184] Before a predefined fixed number of iterations, when the particle individual is a non-dominated point, the particle individual is calculated by the numerical simulation simulation model to update the variable confidence proxy model;

[0185] After a predefined fixed number of iterations, the particle individual with the smallest Euclidean distance from the second production scheme is selected and input into the variable confidence proxy model to update the variable confidence proxy model.

[0186] Based on the Gaussian process regression model theory, the prediction uncertainty at point x 0 is represented as follows:

[0187] (14)

[0188] The prediction bias x ( I 0) at point x 0 (the prediction bias is represented by the difference between the value of the second production scheme predicted by the variable confidence model and the value calculated by the numerical simulation (the previous simulation model)) is within the 95.5% confidence interval of the variable confidence model. According to the 2 σ If the dominance status of a production scheme individual does not change due to the prediction uncertainty of the variable confidence proxy model, the objective function is calculated by the variable confidence proxy model, otherwise it is calculated by model B. The selection criteria for the production scheme individual is the relationship between the minimum distance minimum value (MMD) and the prediction bias (that is, the relationship between MMD and prediction bias, who is larger or smaller, that is, the relationship between MMD and prediction bias, if MMD is greater than prediction bias, the model does not need to be updated, MMD is less than prediction bias, the model needs to be updated), such as Figure 8As shown in the figure below. During the selection process, an individual's MMD is defined as the minimum distance between dominant and non-dominant points. The specific calculation process of MMD is shown in the figure below, representing the distance between points. d Defined as the Euclidean distance between two points.

[0189] Meanwhile, MMD and prediction bias have two different relationships, such as Figure 9 As shown:

[0190] The first type is (a), where prediction bias does not affect the dominance relationship between individuals; the second type is (b), where prediction bias does affect the dominance relationship between individuals. When the first type is met, it is not necessary to use Model B to calculate the data sample; when the second type is met, it is necessary to use Model B to calculate the data sample.

[0191] The formula for selecting Model B to calculate the data sample is as follows:

[0192] (15)

[0193] in:

[0194] (16)

[0195] (2) Population-based model update strategy.

[0196] To ensure the diversity of the search population during the optimization iteration process, a population-based update strategy is considered. After a predefined fixed number of iterations, samples with the minimum Euclidean distance among other HF samples (second production scheme) are selected for HF analysis. In summary, samples meeting the following conditions are selected:

[0197] (17)

[0198] in Xp This is the current high-precision sample set. l This represents the number of samples in the current high-precision sample set. d ( x , x m )for x With the m Euclidean distance between the current high-precision sample sets.

[0199] The two model update strategies mentioned above continuously improve the accuracy of the variable credibility proxy model during the optimization process.

[0200] In each iteration, new samples are added to the variable credibility agent model to improve its prediction accuracy, which is based on the update strategy that considers individuals and the population mentioned above.

[0201] Therefore, the effects of the present invention are as follows:

[0202] 1. This invention establishes reservoir numerical simulation models of different precision to achieve sampling of multi-source, multi-precision data samples and fully leverage the advantages of different data samples.

[0203] 2. This invention utilizes a scaling function-based variable confidence surrogate modeling method to achieve efficient fusion modeling of multi-source, multi-precision reservoir numerical simulation data.

[0204] 3. In the optimization iteration process, this invention considers two different agent model update strategies: individual-based and population-based, to improve the predictive performance of the agent model.

[0205] 4. Based on the integration of multi-source, multi-precision data sample variable credibility proxy modeling, this invention combines a hybrid multi-objective example swarm algorithm to leverage its ability to effectively balance global and local search, thereby achieving multi-objective optimization decision-making for carbonate reservoir production schemes.

[0206] 5. This invention proposes a complete framework for variable confidence proxy optimization of carbonate reservoirs based on multi-source, multi-precision data fusion, encompassing the definition of numerical models with different precisions, the setting of optimization models, the sampling of data samples, the training of variable confidence models, model update strategies, and hybrid multi-objective algorithms. This framework enables rapid response of the objective function during the production scheme optimization iteration process, replacing time-consuming numerical simulation calculations. It also overcomes the problem that traditional proxy models cannot adapt to the complex reservoir characteristics of carbonate reservoirs, greatly improving optimization efficiency and the quality of optimization schemes.

[0207] Taking a single injection and production well in a carbonate rock formation as an example:

[0208] 1. Establishment of numerical model for carbonate reservoirs:

[0209] Following step S1, a grid size of 1000 and a reservoir size are established using a stochastic modeling method. 100 A 100×10 model of a carbonate reservoir considering only porosity, with a grid size of 1000, a fracture count of 80, and a reservoir size of [missing information]. 100 Carbonate reservoirs with porosity and fractures considered simultaneously (×100×10). See examples below. Figure 2 and 3 As shown.

[0210] 2. Following step S2, set the optimization variable space, considering an injection pressure of 25-30 MPa for injection wells and a bottom hole pressure of 10-25 MPa for production wells. Define NPV and production-like quantity as two objective functions, and build a multi-objective optimization model for carbonate reservoir production schemes.

[0211] 3. Multi-source, multi-precision data sampling:

[0212] On the basis of 2 steps, 1000 groups of samples (first production scheme) are generated for A model (low-precision numerical simulation model) sampling and 100 groups of samples (second production scheme) are generated for B model (high-precision numerical simulation model) sampling by using the Latin hypercube sampling method.

[0213] 4. A variable confidence proxy model of multi-source and multi-precision data fusion is established.

[0214] On the data sample set established in step 3, the multi-source and multi-precision data fusion is realized according to step S4, the variable confidence proxy model is constructed, and the accuracy thereof is verified, and the verification effect is as shown in Figure 4 and 5 .

[0215] 5. Multi-objective optimization of production schemes

[0216] According to step F, the advantages of the hybrid multi-objective particle swarm algorithm are exerted, and the model updating iteration strategy based on individuals and populations in step S5 is considered in the optimization iteration process, the further improvement of the prediction performance of the proxy model in the optimization process is realized, and the optimal Pareto solution set is as shown in Figure 6 , and a production scheme with relatively large NPV and oil production is selected and displayed as shown in Figure 7 .

[0217] The above specific embodiments further specifically describe the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method, characterized in that, The application relates to a method for multi-objective optimization of production schemes of carbonate reservoirs. The method comprises the following steps: building a low-precision carbonate reservoir numerical model considering only pores and a high-precision carbonate reservoir numerical model considering both pores and fractures; Based on the design space, the Latin hypercube sampling method is used to sample the optimization variables to generate... One and A first production plan and a second production plan are provided, wherein the first production plan is used to calculate the objective function using the low-precision carbonate reservoir numerical model, and the second production plan is used to calculate the objective function using both the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model. setting optimization variables, a design space and a production scheme multi-objective optimization model with economic parameters and cumulative oil production as target functions; establishing a variable credibility proxy model for fusing target functions calculated by the first production scheme and the second production scheme based on a scaling function; optimizing each parameter in the production scheme multi-objective optimization model based on a hybrid particle swarm algorithm, wherein, in the optimization process, the variable credibility proxy model is updated based on particle individuals and populations in the hybrid particle swarm algorithm; the individual-based variable credibility proxy model updating strategy is as follows: before a predefined fixed iteration number, when a particle individual is a non-dominated point, the particle individual is calculated by a numerical simulation simulation model to update the variable credibility proxy model; after the predefined fixed iteration number, a particle individual with the minimum Euclidean distance from the second production scheme is selected and input into the variable credibility proxy model to update the variable credibility proxy model; the population-based variable credibility proxy model updating strategy is as follows:

2. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 1, characterized in that, The vector form of the optimization variables is represented as where the optimization variables are and , represents the injection rate of the injection well j, represents the bottom hole pressure of the production well p, the design space includes the injection rate range of the injection well and the bottom hole pressure range of the production well, wherein the injection rate range of the injection well is represented as and the bottom hole pressure range of the production well is represented as wherein , represents the lower limit value and the upper limit value of the injection rate of the injection well, respectively, , represents the lower limit value and the upper limit value of the bottom hole pressure range of the production well, respectively.

3. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 2, characterized in that, in order to guarantee the diversity of a scheme search population in the optimization iteration process, the population-based updating strategy in the optimization iteration process is considered, and after the predefined fixed iteration number, samples with the minimum Euclidean distance from the second production scheme are selected for HF analysis. ； the target function is as follows: ； wherein represents an economic parameter, represents the cumulative oil production, g ( u ) is a constraint, lb and the constraint condition is as follows: are lower and upper limit values for the optimization variable, u is a vector form of the optimization variable; ； ； in, For oil prices; The price is for water treatment. Gas processing price; This is the price for water injection; For gas injection price, 、 and For the first j Koujing in the n Oil production, water production, and gas production over a specific time period; and For the first j Koujing in the n Water and gas injection volumes for each time period; This refers to the number of production time steps. b The discount rate; P The number of production wells; I The number of injection wells; For production time steps, For the first n The length of each time step.

4. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 2, characterized in that, Based on the design space, the Latin hypercube sampling method is used to sample the optimization variables to generate... Individual and The specific process of calculating the objective function using the low-precision carbonate reservoir numerical model and the high-precision carbonate reservoir numerical model, based on the first and second production schemes, is as follows: A1, generation one first production plan: A1.

1. dividing the injection volume and the bottom hole pressure respectively into subspaces, thereby causing the design space to be divided into subspaces; A1.

2. In random samples, each of which comprises an injection rate and a bottom hole pressure, thereby obtaining first production scenarios , denotes the i-th first production scenario, ;​ A2, generating A second production scheme is: A2.

1. dividing the injection volume and the bottom hole pressure respectively into equal parts of the design space, thereby causing the design space to be divided into equal parts of the design space; A2.

2. In random samples, each of which comprises an injection rate and a bottom hole pressure, thereby obtaining second production scenarios , denotes the i-th second production scenario, ;​ A3. Using the aforementioned low-precision carbonate reservoir numerical model to... The objective function of the first production scheme is calculated to obtain... The first objective function response value, The response value of the first objective function is expressed as follows: ; A4. Using the high-precision carbonate reservoir numerical model to... The objective function of the second production scheme is calculated to obtain... The response value of the second objective function. The response value of the second objective function is expressed as follows: ; A5. Using the aforementioned low-precision carbonate reservoir numerical model to... The objective function of the second production scheme is calculated to obtain... The response value of the third objective function, The response value of the third objective function is expressed as follows: .

5. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 4, characterized in that, ub ； wherein, represents a variable credibility proxy model, wherein a set of sampled data x includes an injection rate and a bottom hole pressure, represents a scale function characterized by a Gaussian process regression model , , , .

6. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 5, characterized in that, The specific expression is: ； ； is the correlation vector is the correlation matrix, is a column vector of length n h of length 7. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 3, characterized in that, the variable credibility proxy model is as follows: the specific process of optimizing each parameter in the production scheme multi-objective optimization model based on the hybrid particle swarm algorithm is as follows: B1, initializing parameters, setting a maximum iteration number, a maximum speed of particle individuals, a population size N and a termination condition, and initializing positions and speeds of the particle individuals to form a population, wherein the population is composed of a plurality of optimization variables; B3. Selecting a global optimum from the population according to the fitness values and individual optimums ; B2, calculating the fitness value of each particle individual; ； wherein, represents the i-th particle individual in the j-th dimension in the iteration number, represents the i-th particle individual in the j-th dimension in the iteration number, , represents the injection amount, represents the bottom hole pressure, represents the i-th particle individual in the j-th dimension in the iteration number, represents the i-th particle individual in the j-th dimension in the iteration number, wherein, is a scaling factor, and is a random number, N is a positive integer, represents the i-th particle individual in the j-th dimension in the iteration number, represents the i-th particle individual in the j-th dimension in the iteration number, is a crossover probability; B4, randomly selecting half of the particle individuals in the population, performing mutation operation by using a DE / rand-to-best operator, and simultaneously, randomly selecting particle individuals in the population to perform crossover operation according to a set crossover operation operator, and the crossover operation operator is specifically as follows: ； ； ； simultaneously, the other half of the particle individuals in the population are updated according to the following formula: ； wherein, represents the first particle individual after iteration, represents the first particle individual after iteration, represents the first fitness value of the particle individual after iteration, represents the first fitness value of the particle individual after iteration; B6, judging whether the termination condition is met, if yes, outputting the global optimal value , otherwise, jumping to B7; B5, calculating the fitness value of each particle individual in the new population, and updating the population by using a selection operation operator, wherein the selection operation operator is specifically as follows: B7, judging whether a predefined fixed iteration number is reached, wherein the fixed iteration number is less than the maximum iteration number, if yes, jumping to B8, otherwise, returning to B4; B8, taking the minimum Euclidean distance between the particle individuals in the current population and samples in a high-precision sample set composed of the second production scheme except the population as a new particle individual to update the population, and then jumping to B6.

8. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 7, characterized in that, The fitness value of the particle individual is an objective function of the particle individual, and when the fitness value of the particle individual is calculated, if the dominance state of the particle individual is not changed due to the prediction uncertainty of the variable confidence proxy model, the fitness value of the particle individual is calculated by the variable confidence proxy model, otherwise, the fitness value of the particle individual is calculated by the high-precision carbonate reservoir numerical model.

9. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 8, characterized in that, The representation of the prediction uncertainty of the particle individual is as follows: ； wherein is a variance; is a correlation vector of the scaling function; is a correlation matrix of the scaling function; denotes the i-th particle individual, if , and M is the number of particle individuals in the population, then the dominance status of the particle individual is determined not to have changed due to the prediction uncertainty of the variable confidence surrogate model, wherein ; , denotes a dominated point, denotes an non-dominated point, is the minimum of the minimum distances between all dominated points and all non-dominated points, denotes the prediction uncertainty of the particle individual.

10. The multi-source multi-precision data fusion lithologic reservoir variable reliability agent optimization method according to claim 7, characterized in that, In B8, the formula corresponding to population updating is: ； wherein, is a high-precision sample set, l is a number of samples in the high-precision sample set, denotes the Euclidean distance between the particle individual i of the current population and the m-th sample of the high-precision sample set.

Images