CN117236195B - Machine learning offline agent model production optimization method for reducing development risk - Google Patents
Machine learning offline agent model production optimization method for reducing development risk Download PDFInfo
- Publication number
- CN117236195B CN117236195B CN202311490309.9A CN202311490309A CN117236195B CN 117236195 B CN117236195 B CN 117236195B CN 202311490309 A CN202311490309 A CN 202311490309A CN 117236195 B CN117236195 B CN 117236195B
- Authority
- CN
- China
- Prior art keywords
- injection
- population
- production
- individual
- production system
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000004519 manufacturing process Methods 0.000 title claims abstract description 215
- 238000000034 method Methods 0.000 title claims abstract description 161
- 238000005457 optimization Methods 0.000 title claims abstract description 68
- 238000011161 development Methods 0.000 title claims abstract description 42
- 238000010801 machine learning Methods 0.000 title claims abstract description 20
- 238000002347 injection Methods 0.000 claims abstract description 202
- 239000007924 injection Substances 0.000 claims abstract description 202
- 230000035699 permeability Effects 0.000 claims abstract description 120
- 230000006870 function Effects 0.000 claims abstract description 86
- 230000008569 process Effects 0.000 claims abstract description 44
- 238000004088 simulation Methods 0.000 claims abstract description 29
- 239000003795 chemical substances by application Substances 0.000 claims abstract description 28
- 239000013598 vector Substances 0.000 claims description 65
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 35
- 239000011159 matrix material Substances 0.000 claims description 22
- 238000005070 sampling Methods 0.000 claims description 19
- 230000035772 mutation Effects 0.000 claims description 15
- 230000009467 reduction Effects 0.000 claims description 13
- 238000011160 research Methods 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 12
- 230000003204 osmotic effect Effects 0.000 claims description 9
- 238000012847 principal component analysis method Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000006467 substitution reaction Methods 0.000 claims description 2
- 238000000513 principal component analysis Methods 0.000 abstract 1
- 230000001186 cumulative effect Effects 0.000 description 12
- 230000008901 benefit Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000000243 solution Substances 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
- Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping
Abstract
The invention discloses a machine learning offline agent model production optimization method for reducing development risk, and relates to the technical field of oil reservoir production optimization. The method comprises the steps of extracting main characteristics of each permeability field by utilizing principal component analysis, clustering each permeability field, selecting a representative permeability field to construct a plurality of oil reservoir numerical simulation models, taking weighted average net present value of each oil reservoir numerical simulation model as an objective function, establishing an oil reservoir injection and production optimization problem solving model considering uncertainty of the permeability field, constructing an offline database and an initial population, respectively establishing a plurality of radial basis function proxy models and kriging function proxy models by utilizing the offline database, carrying out iterative optimization by taking the self-adaptive selection proxy model as an optimization target, and increasing population diversity by combining multiple offspring strategies in an optimization process to obtain an optimal development scheme. According to the method, the off-line database is fully utilized to guide the injection and production optimization process, the optimization time is shortened, meanwhile, the efficient injection and production scheme of the oil reservoir is accurately obtained, and the scheme risk is reduced.
Description
Technical Field
The invention relates to the technical field of oil reservoir production optimization, in particular to a machine learning offline agent model production optimization method for reducing development risk.
Background
For an oil reservoir with complex conditions and strong heterogeneity, injected water is easy to finger in along a hypertonic channel and limited in the production process, and the injection and production optimization technology realizes the optimal displacement effect by optimizing the production system of a water injection well and a production well, so that the method has become a main means for improving economic benefit.
At present, the existing conventional injection and production optimization method mainly takes the net present value of a single oil reservoir numerical simulation model as a target value, and optimizes the injection and production degree of an operation well to guide the production and development of an oil field. However, the effectiveness of the final scheme obtained by using a single oil reservoir numerical simulation model to represent the block for optimization depends on the parameter accuracy of the oil reservoir numerical simulation model, but in the modeling process of the actual oil reservoir numerical simulation model, the initial model often has larger uncertainty, and the actual situation of the oil reservoir cannot be accurately reflected. Even after the history fitting, the uncertainty can only be reduced but not eliminated. Therefore, the injection and production optimization performed by using only a single model can cause poor development effect and face higher development risk because the difference between the adopted oil reservoir numerical simulation model and the underground actual parameters is too large.
In order to reduce development risk and improve reliability of an injection and production optimization scheme, injection and production optimization can be performed on a plurality of permeability fields obtained after history fitting, but under the method, a large amount of numerical simulation time is consumed to calculate objective function values of an oil reservoir numerical simulation model constructed in all permeability fields, and the actual requirements of oil field development are difficult to meet.
The machine learning agent model based on data driving takes a production system as input, an economic net present value as output, and an agent model with higher calculation speed is trained and constructed to replace an optimization objective function to instantly evaluate a large number of candidate injection and production degrees generated by an evolutionary algorithm, so that evolutionary optimization is guided, and the number of times of real numerical simulation is reduced. The data driven proxy model method can be divided into an online method and an offline method according to whether a numerical simulator is called in the optimization process to evaluate and generate new training data. The off-line method only needs to call the numerical simulator for simulation when the off-line database is constructed, actual evaluation is not needed in the actual optimization process, so that the on-site real-time optimization requirement can be met, the injection and production optimization scheme can be provided immediately, the numerical simulation software can be separated in the optimization process, and the method is suitable for solving the injection and production optimization high-time-consuming problem.
Since the offline method does not evaluate the data for a true value to verify the quality of the proxy model, or to verify the best solution, it is a very critical issue whether the accuracy of the proxy model constructed using the offline database meets the actual optimization requirements. The existing offline method mainly builds a proxy model based on a single machine learning model, and can obtain better performance on a specific oil reservoir block, but is difficult to adapt to the injection and production problems of all oil reservoir block types simultaneously and has a larger improvement space in accuracy. Therefore, a higher-precision offline agent model framework is needed, and the performance of the offline data-driven injection and production optimization method is improved.
Disclosure of Invention
Aiming at the problems that the oil reservoir development risk is high and the accuracy of the existing offline agent model method is low due to the fact that the distribution situation of all permeability fields in the oil reservoir is difficult to comprehensively consider in the prior art, the invention provides a machine learning offline agent model production optimization method for reducing the development risk.
In order to achieve the above purpose, the invention adopts the following technical scheme:
a machine learning offline agent model production optimization method for reducing development risk comprises the following steps:
step 1, acquiring a water injection development mode of a research area, performing history fitting to obtain a permeability field set consisting of a plurality of permeability fields, inputting each permeability field into an oil reservoir numerical simulator, and establishing a plurality of oil reservoir numerical simulation models;
step 2, after main characteristics of each permeability field are extracted by using a principal component analysis method, clustering the permeability fields based on a K-Means method in combination with the extracted main characteristics, randomly selecting a substitution table permeability field from each type of permeability field to obtain a representative oil reservoir, and establishing a representative oil reservoir numerical simulation model set;
step 3, establishing an uncertainty oil reservoir injection and production optimization model of the permeability field by taking the weighted average net present value representing the oil reservoir as an objective function;
step 4, sampling in an injection and production optimization range limited by boundary constraint conditions of an injection and production system based on Latin hypercube sampling methodConstructing a databaseAnd an initial population, wherein an optimal injection and extraction system Xb and a weighted average net present value corresponding to the optimal individuals in the initial population are obtained >；
Step 5, based on Bagging sampling method by using databaseEstablishing T radial basis function proxy models according to the sample data in the model;
step 6, setting the maximum iteration times maxFEs, initializing the iteration times Fes to 0, and starting to perform iterative optimization;
step 7, judging whether the iteration times Fes are larger than the maximum iteration times maxFEs, if the iteration times Fes are larger than the maximum iteration times dimension maxFEs, entering step 10, otherwise, continuously judging whether the iteration times Fes are larger than half of the maximum iteration times dimension maxFEs and whether the iteration times Fes are even; if the iteration times Fes are greater than half of the maximum iteration times dimension maxFEs and the iteration times Fes are even, entering a step 9, otherwise, entering a step 8;
step 8, determining an iteration objective function, carrying out cross mutation on the initial population to form a plurality of groups of offspring population, screening the offspring population by using the iteration objective function, and returning to the step 7 after updating the initial population, the iteration times and the optimal injection and production system;
step 9, constructing a Ke Li jin proxy modelAnd the same population Plocal as the primary population, using the kriging proxy model ++>Carrying out differential evolution on the population Plocal for a plurality of times to update the population Plocal, searching the optimal individuals in the updated initial population by matching with the iteration objective function, and returning to the step 7 after updating the initial population and the iteration times;
And 10, ending iterative optimization, and outputting an optimal development scheme and a weighted average net present value.
Preferably, in the step 2, the method specifically includes the following steps:
step 2.1, obtaining a permeability field set by history fitting to the research area, and obtaining the permeability field setExpanding the two-dimensional permeability fields to obtain one-dimensional vectors, and constructing a permeability field matrix based on each one-dimensional vector>Wherein the size of the permeability field matrix is +.>，/>For the number of two-dimensional permeability fields, +.>Length of one-dimensional vector, +.>，/>For the reservoir numerical simulation model at +.>The number of meshes in the axial direction, +.>For the reservoir numerical simulation model at +.>The number of meshes in the axial direction;
step 2.2, performing dimension reduction treatment on the permeability field matrix based on a principal component analysis method, and compressing a high-dimension permeability field into a low-dimension variable, as shown in a formula (1):
（1）
in the method, in the process of the invention,for permeability field mean +.>Is a feature matrix->For the permeability field characteristic matrix after dimension reduction, < ->Is an original permeability field matrix;
step 2.3, clustering the permeability field matrix after dimension reduction based on the K-Means method, and randomly selectingThe feature vectors are used as cluster centers to obtain +.>Center vectors of the classes are respectively calculated, euclidean distances from other feature vectors to the center vectors of the classes are respectively calculated, the other feature vectors are closely divided into the classes with the nearest Euclidean distances, the center vectors of the classes are updated at any time in the clustering process, and whether the center vectors of the classes are changed or not is judged; if the center vectors of the various types are changed, the Euclidean distance between other feature vectors and the center vectors of the various types is recalculated until the center vectors of the various types are not changed; if the center vectors of the various types are not changed, stopping updating the center vectors of the various types and outputting a clustering result;
And 2.4, dividing the permeability fields in the permeability field set into W types according to the feature similarity of the permeability field matrix after dimension reduction to obtain W permeability field subsets, randomly selecting one permeability field from each permeability field subset as a representative permeability field, taking oil reservoirs corresponding to each permeability field as a representative oil reservoir to form a representative oil reservoir set, establishing an oil reservoir numerical simulation model for each oil reservoir in the representative oil reservoir set, and establishing a representative oil reservoir numerical simulation model set.
Preferably, in the step 3, a weighted average net present value representing the oil reservoir is taken as an objective function, and a permeability field uncertainty oil reservoir injection and production optimization model is established in combination with a boundary constraint condition of an injection and production system, as shown in a formula (2):
（2）
wherein,
（3）
in the method, in the process of the invention,find a function for maximum value +_>Optimizing the objective function for injection and production->For the injection and production system, the recipe is->State variables of reservoir characteristic information calculated for reservoir numerical simulator, < >>Is the real number domain>For the dimension of the injection-production system, the ∈>，/>For the number of wells to be optimized in the reservoir +.>For the total number of time steps>For the lower boundary of the injection-production system +.>The upper boundary of the injection and production system; />The sequence number of the time step; / >Is->The time of each time step is in days; />For the number representing permeability fields in the reservoir set, +.>To represent the total number of reservoirs in the reservoir collection; />Is->Weights of individual reservoirs representing permeability fields in calculating weighted average net present value, +.>，/>Is->The individual reservoirs represent the total number of permeability sites in the classification, < >>The total number of the two-dimensional permeability fields in the permeability field set; />Is the discount rate; />For producing the serial number of the well>The total number of production wells in the target reservoir; />For serial number of injection well>The total number of injection wells in the target reservoir; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily oil production corresponding to the production well, wherein the unit is STB/D; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water yield corresponding to the injection well, wherein the unit is STB/D;is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water injection quantity corresponding to an injection well is shown as STB/D; />The unit is USD/STB; />The cost of water production is treated for unit volume, and the unit is USD/STB; />The cost of water filling per unit volume is USD/STB.
Preferably, in the step 4, the method specifically includes the following steps:
step 4.1, collecting injection and production system samples in an injection and production optimization range limited by boundary constraint conditions of the injection and production system based on Latin hypercube sampling method, and constructing an injection and production system sample set，/>Wherein->For serial number->For the total number of samples->Is->A single injection and production system;
step 4.2, the injection and collection degree sample setThe injection and production degree samples in the process are respectively input into an oil reservoir numerical simulator, weighted average net present values corresponding to the injection and production degree samples are calculated by the oil reservoir numerical simulator, and a weighted average net present value data set is constructed>，/>，/>Is->Individual note picking degree->Is determined by the weighted average net present value of (2);
step 4.3, based on the injection and production degree sample setAnd a weighted average net present value dataset +.>Construction of a database->The database comprises a plurality of sample data;
the database is provided withWherein->Is database->The%>Sample data;
step 4.4, the database is subjected to the order of weighted average net present value from high to lowThe sample data in the initial population is sequenced, the initial population is built by utilizing the first N sample data, the injection and collection degree corresponding to the optimal individual in the initial population is obtained, and the optimal injection and collection degree Xb and the weighted average net present value are obtained >。
Preferably, in the step 5, the method specifically includes the following steps:
step 5.1, based on Bagging sampling method by multiple times from databaseThe method comprises the steps of selecting a plurality of sample data by replacing sampling, randomly constructing training sets, and generating T training sets;
step 5.2, establishing radial basis function proxy models for predicting weighted average net present values corresponding to the injection and production systems aiming at each training set to obtain T radial basis function proxy models;
the radial basis function proxy model is shown in formula (4):
（4）
in the method, in the process of the invention,injection and production degree for radial basis function proxy model>Is a predicted value of (2); />For serial number->Is->Weight of the individual sample data, +.>Is->The system of injection and production is->Is European norm, ++>As a basis function +.>Is the total number of individuals in the initial sample.
Preferably, in the step 8, the method specifically includes the following steps:
step 8.1, respectively calculating predicted values of injection and production systems corresponding to optimal individuals in an initial population by using T radial basis function proxy models, and grouping the radial basis function proxy models in pairs after the radial basis function proxy models are ordered according to the sequence from the large predicted values to the small predicted values;
step 8.2, selecting a radial basis function proxy model from each group to establish a target proxy model set, wherein the target proxy model set contains And the radial basis function proxy models are used for calculating the average value of all the radial basis function proxy model predicted values in the target proxy model set and serving as an iterative target function, and the average value is shown in a formula (5):
（5）
in the method, in the process of the invention,for an iterative objective function determined based on a radial basis function proxy model +.>For the +.>Injection and production system calculated by radial basis function proxy model>Is a predicted value of (2);
step 8.3, randomly generating G groups of hyper-parameter data consisting of variation factors F and crossover factors CR to form a hyper-parameter set，/>Wherein->Is a superparameter set +.>The%>Group superparameter data,/->Is->Variation factor in group super parameter data, +.>Is->Cross factors in the group hyper-parameter data;
step 8.4, performing cross mutation on the initial population by utilizing each super parameter data in the super parameter set to generate G sub-generation populations，/>}, wherein->For utilizing hyper-parameter set->Middle->Group super parameter data control generation +.>A population of offspring;
the cross mutation process is as follows:
（6）
（7）
in the method, in the process of the invention,is->The first->The variation vector of the injection and production system corresponding to each offspring individual,/->Is->The first->The +.about.th of the mutation vector corresponding to each offspring individual >Dimension; />、/>、/>The method comprises the steps of selecting three mutually different offspring individuals from an initial population, and corresponding injection and production degrees; />Progeny population obtained for crossover mutation>Middle->The +.f. of the injection and production system corresponding to each offspring individual>Dimension; />Is the>First digit of random vector>Dimension; />An index selected randomly; />Is>Individual note picking degree->Is>Dimension;
step 8.5, calculating the predicted value of the injection and production system corresponding to each child individual in the G child population by using the iterative objective function;
8.6, comparing the predicted value of the injection and production system corresponding to each child individual in the G child population with the weighted average net present value corresponding to the injection and production system corresponding to the initial population one by one according to the child population numbering sequence, and when the predicted value of the injection and production system corresponding to the child individual in the child population is larger than the weighted average net present value of the injection and production system corresponding to the individual in the initial population, replacing the individual in the initial population by the child individual in the child population to update the initial population;
and 8.7, determining the optimal individuals in the updated initial population, acquiring an optimal injection and production system Xb and a weighted average net present value corresponding to the optimal individuals, updating the optimal injection and production system Xb and iteration times, and returning to the step 7.
Preferably, in the step 9, the method specifically includes the following steps:
step 9.1, respectively calculating the optimal individuals in the initial population and the databaseThe Euclidean distance between the sample data in the database is increased from small to large>The sample data in the sequence are ordered, the first A sample data in the sequence are selected, and a kriging proxy model is constructed>As shown in formula (8):
（8）
in the method, in the process of the invention,for the kriging proxy model->Degree of contrast and harvest->Predicted value of +.>For the average constant of the sample data, +.>Is an autocorrelation error term with zero mean;
step 9.2, constructing a population Plocal which is identical to the initial population, and carrying out differential evolution treatment on the population Plocal to generate a offspring population，/>Wherein->For offspring population->The%>A child generation individual;
the differential evolution processing process is as follows:
（9）
（10）
in the method, in the process of the invention,for offspring population->Middle->The variation vector of the injection and production system corresponding to each individual, < >>For offspring population->Middle->The +.about.th of the variation vector corresponding to each individual>Dimension; />Is a variation factor with a value range of [0.4,1 ]]；/>、/>、/>The injection and production rates corresponding to three mutually different individuals in the population Plocal; />For offspring population->Middle->The +.f. of the injection and production system corresponding to each offspring individual >Dimension; />Is the>First digit of random vector>Dimension; />Is constant, between 0 and 1; />An index selected randomly;
step 9.3, performing differential evolution processing on the population Plocal for a plurality of times according to the preset differential evolution processing times, wherein in each differential evolution processing process, a kriging proxy model is utilizedCalculating offspring population->Predictive value of injection and production system corresponding to each offspring individual, current offspring population +.>When the predicted value of the injection and production system corresponding to the child individuals in the population Plocal is smaller than the weighted average net present value of the injection and production system corresponding to the corresponding individuals in the population Plocal, the child population +.>The offspring individuals in (a) replace individuals in the population Plocal, and the population Ploc is updatedal；
Step 9.4, searching the population Plocal after the differential evolution treatment by using an iterative objective function, and determining the optimal individual in the population Plocal after the differential evolution treatmentCalculating optimal individuals using iterative objective functions>The predicted value of the corresponding injection and production system is weighted average net present value of the optimal injection and production degree Xb corresponding to the optimal individual in the initial population>Comparing; if the optimal individuals in the population Plocal +.>The predicted value of the corresponding injection and production system is smaller than the weighted average net present value of the optimal injection and production system Xb corresponding to the optimal individual in the initial population >Then use the optimal individuals in the population Plocal +.>And (3) replacing the optimal individuals in the initial population, returning to the step (7) after updating the initial population and the iteration times, otherwise, directly updating the iteration times, and returning to the step (7).
The invention has the beneficial effects that:
the invention provides a machine learning offline agent model production optimization method for reducing development risk, which solves the problem that an oil reservoir injection and production optimization scheme is difficult to obtain because the oil reservoir permeation field distribution is difficult to determine. According to the method, the dimension reduction clustering is carried out on the permeability field set obtained based on the water injection development mode history fitting of the research area, and then the representative permeability field is selected in a clustering mode to establish the representative oil reservoir numerical simulation model, so that all possible conditions of the permeability field in the oil reservoir are comprehensively considered, and the injection and production optimization is carried out on each permeability field, so that the development risk of an oil reservoir development scheme is reduced, and redundant calculation in the process of formulating the oil reservoir development scheme is avoided.
Meanwhile, in order to improve the accuracy of the agent model in offline optimization, a high-precision machine learning offline agent model production optimization method is constructed, an offline database is constructed based on Latin hypercube sampling, and the self-adaptive selection agent model strategy and the high-precision agent model strategy are combined for alternate searching, so that the exploration and development performances of the machine learning offline agent model production optimization method are effectively balanced, the accuracy of an oil reservoir optimal development scheme determined by the machine learning offline agent model production optimization method is improved, the optimization speed of optimization calculation is improved, the accurate acquisition of a low-risk high-efficiency injection and production optimization scheme suitable for an oil reservoir with uncertain permeability is realized, the basis is provided for guiding the formulation of a complex reservoir development scheme, and the method has extremely high popularization and application values.
Drawings
FIG. 1 is a flow chart of a machine learning offline proxy model production optimization method for reducing development risk according to the present invention.
FIG. 2 is a flow chart of determining an optimal injection and production regime in the method of the present invention.
FIG. 3 is a plot of weighted average net present value versus iteration number during iterative optimization.
FIG. 4 is a graph of the cumulative oil production weighted average for an optimal injection and production scheme determined using the method of the present invention and a single agent method.
FIG. 5 is a graph of the cumulative weighted average of water production for an optimal injection and production scheme determined using the method of the present invention and a single agent method.
FIG. 6 is a graph of the cumulative water injection weighted average for an optimal injection and production scheme determined using the methods of the present invention and a single agent method.
Detailed Description
The invention is described in further detail below with reference to the attached drawings and detailed description:
the invention provides a machine learning offline agent model production optimization method for reducing development risk, as shown in fig. 1, a study block in the embodiment adopts a water injection development mode, 4 water injection wells and 9 production wells are arranged in the study block, and the method is used for determining the optimal development scheme of the study block and specifically comprises the following steps:
Step 1, selecting a research area, acquiring a water injection development mode of the research area, dividing the research area into grids, and arranging the research area in the following wayDividing into 60 grids in the axial direction, at +.>Grid divided into 60 pieces in axial direction, grid size is +.>And performing history fitting on a research area to obtain 100 conditions of possible distribution of permeability fields, forming a permeability field set, inputting permeability field parameters corresponding to each permeability field in the permeability field set into an oil reservoir numerical simulator Eclipse, and establishing a plurality of oil reservoir numerical simulation models with the same parameters except for different permeability fields.
Step 2, for a permeability field set, extracting main features of each permeability field in the permeability field set by using a principal component analysis method, clustering each permeability field in the permeability field set based on a K-Means method in combination with the extracted main features, randomly selecting one permeability field from each type of permeability field as a representative permeability field, taking oil reservoirs corresponding to each permeability field as a representative oil reservoir, forming a representative oil reservoir set, establishing an oil reservoir numerical simulation model for each oil reservoir in the representative oil reservoir set, and establishing a representative oil reservoir numerical simulation model set, wherein the method specifically comprises the following steps:
Step 2.1, expanding 100 two-dimensional permeability fields in a permeability field set to obtain one-dimensional vectors, and constructing a permeability field matrix based on each one-dimensional vector, wherein the size of the permeability field matrix is as follows。
In the step 2.2 of the method,performing dimension reduction treatment on the permeability field matrix based on a principal component analysis method, compressing a high-dimension permeability field into a low-dimension variable to obtain a dimension-reduced permeability field characteristic matrixAs shown in formula (1):
（1）
in the method, in the process of the invention,for permeability field mean +.>Is a feature matrix->For the permeability field characteristic matrix after dimension reduction, < ->Is the original permeability field matrix.
Step 2.3, clustering the permeability field matrix after dimension reduction based on a K-Means method, randomly selecting five feature vectors as clustering centers to obtain five types of center vectors, respectively calculating Euclidean distances from other feature vectors to various types of center vectors, dividing other feature vectors into the types with the nearest Euclidean distances nearby, updating the various types of center vectors at any time in the clustering process, and judging whether the various types of center vectors are changed or not; if the center vectors of the various types are changed, the Euclidean distance between other feature vectors and the center vectors of the various types is recalculated until the center vectors of the various types are not changed; if the center vectors of the various types are not changed, stopping updating the center vectors of the various types and outputting a clustering result.
And 2.4, dividing the permeability fields in the permeability field set into five types according to the feature similarity of the permeability field matrix after dimension reduction, obtaining five permeability field subsets, randomly selecting one permeability field from each permeability field subset as a representative permeability field, taking oil reservoirs corresponding to each permeability field as a representative oil reservoir, forming a representative oil reservoir set, establishing an oil reservoir numerical simulation model for each oil reservoir in the representative oil reservoir set, and establishing a representative oil reservoir numerical simulation model set.
And 3, establishing a permeability field uncertainty oil reservoir injection and production optimization model by taking a weighted average net present value representing the oil reservoir as an objective function and combining boundary constraint conditions of an injection and production system, wherein the model is shown in a formula (2):
（2）
wherein,
（3）
in the method, in the process of the invention,find a function for maximum value +_>Optimizing the objective function for injection and production->For the injection and production system, the recipe is->State variables of reservoir characteristic information calculated for reservoir numerical simulator, < >>Is the real number domain>For the dimension of the injection-production system, the ∈>，/>For the number of wells to be optimized in the reservoir +.>For the total number of time steps>For the lower boundary of the injection-production system +.>The upper boundary of the injection and production system; />The sequence number of the time step; />Is->The time of each time step is in days; / >For the number representing permeability fields in the reservoir set, +.>To represent the total number of reservoirs in the reservoir collection; />Is->Weights of individual reservoirs representing permeability fields in calculating weighted average net present value, +.>，/>Is->The individual reservoirs represent the total number of permeability sites in the classification, < >>The total number of the two-dimensional permeability fields in the permeability field set; />Is the discount rate; />For producing the serial number of the well>The total number of production wells in the target reservoir; />For serial number of injection well>The total number of injection wells in the target reservoir; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily oil production corresponding to the production well, wherein the unit is STB/D; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water yield corresponding to the injection well, wherein the unit is STB/D;is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water injection quantity corresponding to an injection well is shown as STB/D; />The unit is USD/STB; />The cost of water production is treated for unit volume, and the unit is USD/STB; />The cost of water filling per unit volume is USD/STB.
In this embodiment, the number of injection and production wells to be optimized in this embodiment is 13, and the injection and production degree to be optimized For the production amount of 9 production wells and the water injection amount of 4 water injection wells, the injection and production optimization objective function is +.>Is a weighted average net present value. In this embodiment, the total number of time steps +.>Is set to 10, the interval between the adjacent time steps is set to 180 days, the production period is 1800 days, and the unit volume of oil production income is +.>Set to 20, singlyBit volume treatment water production cost->Set to 3, water injection cost per unit volume +.>Set to 1, discount rate->Set to 0. According to the number of injection and production wells to be optimized in the embodiment, the dimension of the injection and production degree is determined +.>130, wherein the lower limit of the injection and production rate of 4 water injection wells to be optimized is 145m 3 The upper boundary is set to 290m 3 The method comprises the steps of carrying out a first treatment on the surface of the The lower limit of the injection and production degree of 9 production wells is set to 0m 3 The upper boundary is set to 70m 3 。
Step 4, sampling in the injection and production optimization range limited by the boundary constraint condition of the injection and production system based on Latin hypercube sampling method, obtaining injection and production degree samples, calculating weighted average net present value of each injection and production degree sample by using a numerical simulator, and constructing a databaseAnd in the database->Selecting sample data to establish an initial population, and obtaining an optimal injection and extraction system Xb and a weighted average net present value corresponding to an optimal individual in the initial population >The method specifically comprises the following steps:
step 4.1, collecting injection and production system samples in an injection and production optimization range limited by boundary constraint conditions of the injection and production system based on Latin hypercube sampling method, and constructing an injection and production system sample set，/>Wherein->For serial number->For the total number of samples->Is->A single injection and production system; the total number of samples in this example is set to 1000.
Step 4.2, the injection and collection degree sample setThe injection and production degree samples in the process are respectively input into an oil reservoir numerical simulator, weighted average net present values corresponding to the injection and production degree samples are calculated by the oil reservoir numerical simulator, and a weighted average net present value data set is constructed>，/>，/>Is->Individual note picking degree->To a weighted average net present value of (c).
Step 4.3, based on the injection and production degree sample setAnd a weighted average net present value dataset +.>Construction of a database->The database comprises a plurality of sample data;
the database is provided withWherein->Is database->The%>And sample data.
Step 4.4, the database is subjected to the order of weighted average net present value from high to lowThe sample data in the initial population is sequenced, the initial population is established by utilizing the first N sample data, the optimal individuals in the initial population are obtained, and the optimal injection and collection degree Xb and the corresponding net present value +. >The size N of the initial population in this example is set to 100.
Step 5, based on Bagging and sampling method, database is sampledThe sample data in the model are divided into T training sets, radial basis function proxy models are respectively established for the training sets, and T radial basis function proxy models are obtained, and the method specifically comprises the following steps:
step 5.1, based on Bagging sampling method by multiple times from databaseThe method comprises the steps of sampling a plurality of sample data, randomly constructing a training set, and generatingThe number T of training sets in this embodiment is set to 100.
Step 5.2, establishing radial basis function proxy models for predicting weighted average net present values corresponding to the injection and production systems aiming at each training set to obtain T radial basis function proxy models;
the radial basis function proxy model is shown in formula (4):
（4）
in the method, in the process of the invention,injection and production degree for radial basis function proxy model>Is a predicted value of (2); />For serial number->Is->Weight of the individual sample data, +.>Is->The system of injection and production is->Is European norm, ++>As a basis function +.>Is the total number of individuals in the initial sample.
And 6, setting the maximum iteration times maxFEs as 500 times, initializing the iteration times Fes as 0, combining the initial population based on the Kerling proxy model and the radial basis function proxy model, and entering a step 8 to perform iterative optimization, as shown in fig. 2.
Step 7, judging whether the iteration times Fes are larger than the maximum iteration times maxFEs, if the iteration times Fes are larger than the maximum iteration times dimension maxFEs, entering step 10, otherwise, continuously judging whether the iteration times Fes are larger than half of the maximum iteration times dimension maxFEs and whether the iteration times Fes are even; if the iteration times Fes are greater than half of the maximum iteration times dimension maxFEs and the iteration times Fes are even, entering a step 9, otherwise, entering a step 8.
Step 8, determining an iteration objective function, carrying out cross mutation on an initial population to form a plurality of groups of offspring populations, screening the offspring populations by using the iteration objective function, and returning to step 7 after updating the initial population, the iteration times and the optimal injection and production system, wherein the method specifically comprises the following steps:
and 8.1, respectively calculating predicted values of injection and production systems corresponding to the optimal individuals in the initial population by using T radial basis function proxy models, and grouping the radial basis function proxy models in pairs after the radial basis function proxy models are ordered according to the sequence from the predicted values to the low values.
Step 8.2, selecting a radial basis function proxy model from each group to establish a target proxy model set, wherein the target proxy model set contains And the radial basis function proxy models are used for calculating the average value of all the radial basis function proxy model predicted values in the target proxy model set and serving as an iterative target function, and the average value is shown in a formula (5):
（5）
in the method, in the process of the invention,is based on radial basis function proxyIterative objective function determined by a model, +.>For the +.>Injection and production system calculated by radial basis function proxy model>Is a predicted value of (a).
Step 8.3, randomly generating G groups of hyper-parameter data consisting of variation factors F and crossover factors CR to form a hyper-parameter set，/>Wherein->Is a superparameter set +.>The%>Group superparameter data,/->Is->Variation factor in group super parameter data, +.>Is->The cross factor in the group hyper-parameter data.
Step 8.4, performing cross mutation on the initial population by utilizing each super parameter data in the super parameter set to generate G sub-generation populations，/>}, wherein->For utilizing hyper-parameter set->Middle->Group super parameter data control generation +.>The number G of offspring populations is set to 10 in this example.
The cross mutation process is as follows:
（6）
（7）
in the method, in the process of the invention,is->The first->The variation vector of the injection and production system corresponding to each offspring individual,/- >Is->The first->The +.about.th of the mutation vector corresponding to each offspring individual>Dimension; />、/>、/>The method comprises the steps of selecting three mutually different offspring individuals from an initial population, and corresponding injection and production degrees; />Progeny population obtained for crossover mutation>Middle->The +.f. of the injection and production system corresponding to each offspring individual>Dimension; />Is the>First digit of random vector>Dimension; />An index selected randomly; />Is>Individual note picking degree->Is>Dimension.
And 8.5, calculating the predicted value of the injection and production system corresponding to each child individual in the 10 child population by using the iterative objective function.
And 8.6, comparing the predicted value of the injection and production system corresponding to each child individual in the child population with the weighted average net present value corresponding to the optimal injection and production system in the initial population, and when the predicted value of the injection and production system corresponding to the child individual in the child population is larger than the weighted average net present value of the injection and production system corresponding to the corresponding individual in the initial population, replacing the individual in the initial population by the child individual in the child population, and updating the initial population.
And 8.7, determining the optimal individuals in the updated initial population, acquiring an optimal injection and production system Xb and a weighted average net present value corresponding to the optimal individuals, updating the optimal injection and production system Xb and iteration times, and returning to the step 7.
Step 9, constructing a Ke Li jin proxy modelAnd the same population Plocal as the primary population, using the kriging proxy model ++>And (3) carrying out differential evolution on the population Plocal for a plurality of times to update the population Plocal, searching the optimal individuals in the updated initial population by matching with the iterative objective function, and returning to the step (7) after updating the initial population and the iteration times, wherein the method specifically comprises the following steps of:
step 9.1, respectively calculating the optimal individuals in the initial population and the databaseEuclidean distance between each sample data, and according to euclidean distanceOrder from small to large for database +.>The sample data in the sequence are ordered, the first 250 sample data in the sequence are selected, and a kriging proxy model is constructed>As shown in formula (8):
（8）
in the method, in the process of the invention,for the kriging proxy model->Degree of contrast and harvest->Predicted value of +.>For the average constant of the sample data, +.>Is an autocorrelation error term with zero mean.
Step 9.2, constructing a population Plocal which is identical to the initial population, and carrying out differential evolution treatment on the population Plocal to generate a offspring population，/>Wherein->For offspring population->The%>A child generation individual;
the differential evolution processing process is as follows:
（9）
（10）
in the method, in the process of the invention,for offspring population->Middle- >The variation vector of the injection and production system corresponding to each individual, < >>For offspring population->Middle->The +.about.th of the variation vector corresponding to each individual>Dimension; />Is a variation factor with a value range of [0.4,1 ]]；/>、/>、/>The injection and production rates corresponding to three mutually different individuals in the population Plocal; />For offspring population->Middle->The +.f. of the injection and production system corresponding to each offspring individual>Dimension; />Is the>First digit of random vector>Dimension; />Is constant, between 0 and 1; />Is a randomly selected index.
Step 9.3, performing differential evolution processing on the population Plocal for a plurality of times according to the preset differential evolution processing times, wherein in each differential evolution processing process, a kriging proxy model is utilizedCalculating offspring population->Predictive value of injection and production system corresponding to each offspring individual, current offspring population +.>Neutrons (neutrons)When the predicted value of the injection and production system corresponding to the generation individuals is smaller than the weighted average net present value of the injection and production system corresponding to the corresponding individuals in the population Plocal, the offspring population +.>The offspring individuals in (a) replace individuals in the population Plocal, and the population Plocal is updated. />
Step 9.4, searching the population Plocal after the differential evolution treatment by using an iterative objective function, and determining the optimal individual in the population Plocal after the differential evolution treatment Calculating optimal individuals using iterative objective functions>The predicted value of the corresponding injection and production system is weighted average net present value of the optimal injection and production degree Xb corresponding to the optimal individual in the initial population>Comparing; if the optimal individuals in the population Plocal +.>The predicted value of the corresponding injection and production system is smaller than the weighted average net present value of the optimal injection and production system Xb corresponding to the optimal individual in the initial population>Then use the optimal individuals in the population Plocal +.>And (3) replacing the optimal individuals in the initial population, returning to the step (7) after updating the initial population and the iteration times, otherwise, directly updating the iteration times, and returning to the step (7).
And 10, ending iterative optimization, and outputting an optimal development scheme and a weighted average net present value.
In order to verify the effect of the optimal development scheme determined by the method, an off-line single-agent method and the method for optimizing the research block by using the traditional radial basis function-assisted differential evolution algorithm are respectively utilized to optimize and determine the optimal injection and production system of the research block, namely the optimal development scheme of the research block, so as to obtain a change curve of a weighted average net present value along with the iteration times in the iterative optimization process, as shown in fig. 3. As can be seen from fig. 3, the best development scheme obtained by the method of the present invention has higher economic benefit than the best development scheme obtained by the single agent method.
The cumulative oil production weighted average value, the cumulative water production weighted average value and the cumulative water injection weighted average value of the optimal injection and production scheme determined by the method of the invention and the single agent method are respectively compared, wherein, fig. 4 is a change chart of the cumulative oil production weighted average value of the optimal injection and production scheme determined by the method of the invention and the single agent method, fig. 5 is a change chart of the cumulative water production weighted average value of the optimal injection and production scheme determined by the method of the invention and the single agent method, and fig. 6 is a change chart of the cumulative water injection weighted average value of the optimal injection and production scheme determined by the method of the invention and the single agent method, wherein, the cumulative oil production is a benefit item, the cumulative water production and the cumulative water injection are all cost items. The optimal development scheme determined by the method can improve the accumulated oil yield, reduce the investment cost in the aspects of accumulated water yield and accumulated water injection and have higher economic benefit through analyzing the figures 4-6.
The foregoing description is, of course, merely illustrative of preferred embodiments of the present invention, and it should be understood that the present invention is not limited to the above-described embodiments, but is intended to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
Claims (6)
1. The machine learning offline agent model production optimization method for reducing development risk is characterized by comprising the following steps:
step 1, acquiring a water injection development mode of a research area, performing history fitting to obtain a permeability field set consisting of a plurality of permeability fields, inputting each permeability field into an oil reservoir numerical simulator, and establishing a plurality of oil reservoir numerical simulation models;
step 2, after main characteristics of each permeability field are extracted by using a principal component analysis method, clustering the permeability fields based on a K-Means method in combination with the extracted main characteristics, randomly selecting a substitution table permeability field from each type of permeability field to obtain a representative oil reservoir, and establishing a representative oil reservoir numerical simulation model set;
step 3, establishing an uncertainty oil reservoir injection and production optimization model of the permeability field by taking the weighted average net present value representing the oil reservoir as an objective function;
step 4, sampling in an injection and production optimization range limited by boundary constraint conditions of an injection and production system based on Latin hypercube sampling method, and constructing a databaseAnd an initial population, wherein an optimal injection and extraction system Xb and a weighted average net present value corresponding to the optimal individuals in the initial population are obtained>；
Step 5, based on Bagging sampling method by using database Establishing T radial basis function proxy models according to the sample data in the model;
step 6, setting the maximum iteration times maxFEs, initializing the iteration times Fes to 0, and starting to perform iterative optimization;
step 7, judging whether the iteration times Fes are larger than the maximum iteration times maxFEs, if the iteration times Fes are larger than the maximum iteration times dimension maxFEs, entering step 10, otherwise, continuously judging whether the iteration times Fes are larger than half of the maximum iteration times dimension maxFEs and whether the iteration times Fes are even; if the iteration times Fes are greater than half of the maximum iteration times dimension maxFEs and the iteration times Fes are even, entering a step 9, otherwise, entering a step 8;
step 8, determining an iteration objective function, carrying out cross mutation on the initial population to form a plurality of groups of offspring population, screening the offspring population by using the iteration objective function, and returning to the step 7 after updating the initial population, the iteration times and the optimal injection and production system;
step 9, constructing a Ke Li jin proxy modelAnd the same population Plocal as the primary population, using the kriging proxy model ++>Carrying out differential evolution on the population Plocal for a plurality of times to update the population Plocal, searching the optimal individuals in the updated initial population by matching with the iteration objective function, and returning to the step 7 after updating the initial population and the iteration times;
Step 10, ending iterative optimization, and outputting an optimal development scheme and a weighted average net present value;
in the step 3, a weighted average net present value representing the oil reservoir is taken as an objective function, and a permeability field uncertainty oil reservoir injection and production optimization model is established by combining boundary constraint conditions of an injection and production system, as shown in a formula (2):
（2）
wherein,
（3）
in the method, in the process of the invention,find a function for maximum value +_>Optimizing the objective function for injection and production->For the injection and production system, the recipe is->State variables of reservoir characteristic information calculated for reservoir numerical simulator, < >>Is the real number domain>For the dimension of the injection-production system, the ∈>，/>For the number of wells to be optimized in the reservoir +.>For the total number of time steps>For the lower boundary of the injection-production system +.>The upper boundary of the injection and production system; />The sequence number of the time step; />Is->The time of each time step is in days; />For the number representing permeability fields in the reservoir set, +.>To represent the total number of reservoirs in the reservoir collection; />Is->Weights of individual reservoirs representing permeability fields in calculating weighted average net present value, +.>，/>Is->The individual reservoirs represent the total number of permeability sites in the classification, < >>The total number of the two-dimensional permeability fields in the permeability field set; />Is the discount rate; />For producing the serial number of the well >The total number of production wells in the target reservoir; />For serial number of injection well>The total number of injection wells in the target reservoir; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily oil production corresponding to the production well, wherein the unit is STB/D; />Is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water yield corresponding to the injection well, wherein the unit is STB/D;is->The individual reservoirs represent the osmotic fields at +.>The sequence number of each time step is->Daily water injection quantity corresponding to an injection well is shown as STB/D; />The unit is USD/STB; />The cost of water production is treated for unit volume, and the unit is USD/STB; />The cost of water filling per unit volume is USD/STB.
2. The machine learning offline agent model production optimization method for reducing development risk according to claim 1, wherein in step 2, the method specifically comprises the following steps:
step 2.1, obtaining a permeability field set by history fitting to the research area, and obtaining the permeability field setExpanding the two-dimensional permeability fields to obtain one-dimensional vectors, and constructing a permeability field matrix based on each one-dimensional vector>Wherein the size of the permeability field matrix is ，/>For the number of two-dimensional permeability fields, +.>Length of one-dimensional vector, +.>，/>For the reservoir numerical simulation model at +.>The number of meshes in the axial direction, +.>For the reservoir numerical simulation model at +.>The number of meshes in the axial direction;
step 2.2, performing dimension reduction treatment on the permeability field matrix based on a principal component analysis method, and compressing a high-dimension permeability field into a low-dimension variable, as shown in a formula (1):
（1）
in the method, in the process of the invention,for permeability field mean +.>Is a feature matrix->For the permeability field characteristic matrix after dimension reduction, < ->Is an original permeability field matrix;
step 2.3, clustering the permeability field matrix after dimension reduction based on the K-Means method, and randomly selectingThe feature vectors are used as cluster centers to obtain +.>Center vectors of the individual classes are respectively calculated, euclidean distances from other feature vectors to the center vectors of the various classes are respectively calculated, the other feature vectors are closely divided into the classes with the nearest Euclidean distances, the center vectors of the various classes are updated at any time in the clustering process, and whether the center vectors of the various classes occur is judgedChanging; if the center vectors of the various types are changed, the Euclidean distance between other feature vectors and the center vectors of the various types is recalculated until the center vectors of the various types are not changed; if the center vectors of the various types are not changed, stopping updating the center vectors of the various types and outputting a clustering result;
And 2.4, dividing the permeability fields in the permeability field set into W types according to the feature similarity of the permeability field matrix after dimension reduction to obtain W permeability field subsets, randomly selecting one permeability field from each permeability field subset as a representative permeability field, taking oil reservoirs corresponding to each permeability field as a representative oil reservoir to form a representative oil reservoir set, establishing an oil reservoir numerical simulation model for each oil reservoir in the representative oil reservoir set, and establishing a representative oil reservoir numerical simulation model set.
3. The machine learning offline proxy model production optimization method for reducing development risk according to claim 1, wherein in step 4, the method specifically comprises the following steps:
step 4.1, collecting injection and production system samples in an injection and production optimization range limited by boundary constraint conditions of the injection and production system based on Latin hypercube sampling method, and constructing an injection and production system sample set，/>Wherein->For serial number->For the total number of samples->Is->A single injection and production system;
step 4.2, the injection and collection degree sample setThe injection and production degree samples in the process are respectively input into an oil reservoir numerical simulator, weighted average net present values corresponding to the injection and production degree samples are calculated by the oil reservoir numerical simulator, and a weighted average net present value data set is constructed >，/>，/>Is->Individual note picking degree->Is determined by the weighted average net present value of (2);
step 4.3, based on the injection and production degree sample setAnd a weighted average net present value dataset +.>Construction of a database->The database comprises a plurality of sample data;
the database is provided withWherein->Is database->The%>Sample data;
step 4.4, the database is subjected to the order of weighted average net present value from high to lowThe sample data in the initial population is sequenced, the initial population is built by utilizing the first N sample data, the injection and collection degree corresponding to the optimal individual in the initial population is obtained, and the optimal injection and collection degree Xb and the weighted average net present value are obtained>。
4. The machine learning offline proxy model production optimization method for reducing development risk according to claim 1, wherein in step 5, the method specifically comprises the following steps:
step 5.1, based on Bagging sampling method by multiple times from databaseThe method comprises the steps of selecting a plurality of sample data by replacing sampling, randomly constructing training sets, and generating T training sets;
step 5.2, establishing radial basis function proxy models for predicting weighted average net present values corresponding to the injection and production systems aiming at each training set to obtain T radial basis function proxy models;
The radial basis function proxy model is shown in formula (4):
（4）
in the method, in the process of the invention,injection and production degree for radial basis function proxy model>Is a predicted value of (2); />For serial number->Is->Weight of the individual sample data, +.>Is->The system of injection and production is->Is European norm, ++>As a basis function +.>Is the total number of individuals in the initial sample.
5. The machine learning offline proxy model production optimization method for reducing development risk according to claim 1, wherein in step 8, the method specifically comprises the following steps:
step 8.1, respectively calculating predicted values of injection and production systems corresponding to optimal individuals in an initial population by using T radial basis function proxy models, and grouping the radial basis function proxy models in pairs after the radial basis function proxy models are ordered according to the sequence from the large predicted values to the small predicted values;
step 8.2, selecting a radial basis function proxy model from each group to establish a target proxy model set, wherein the target proxy model set containsAnd the radial basis function proxy models are used for calculating the average value of all the radial basis function proxy model predicted values in the target proxy model set and serving as an iterative target function, and the average value is shown in a formula (5):
（5）
in the method, in the process of the invention, For an iterative objective function determined based on a radial basis function proxy model +.>For the +.>Injection and production system calculated by radial basis function proxy model>Is a predicted value of (2);
step 8.3, randomly generating G groups of hyper-parameter data consisting of variation factors F and crossover factors CR to form a hyper-parameter set，/>Wherein->Is a superparameter set +.>The%>Group superparameter data,/->Is->Variation factor in group super parameter data, +.>Is->Cross factors in the group hyper-parameter data;
step 8.4, performing cross mutation on the initial population by utilizing each super parameter data in the super parameter set to generate G sub-generation populations，/>}, wherein->For utilizing hyper-parameter set->Middle->Group super parameter data control generation +.>A population of offspring;
the cross mutation process is as follows:
（6）
（7）
in the method, in the process of the invention,is->The first->The variation vector of the injection and production system corresponding to each offspring individual,/->Is->The first->The +.about.th of the mutation vector corresponding to each offspring individual>Dimension; />、/>、/>The method comprises the steps of selecting three mutually different offspring individuals from an initial population, and corresponding injection and production degrees; />Progeny population obtained for crossover mutation>Middle- >The +.f. of the injection and production system corresponding to each offspring individual>Dimension; />Is the>First digit of random vector>Dimension; />An index selected randomly; />Is>Individual note picking degree->Is>Dimension;
step 8.5, calculating the predicted value of the injection and production system corresponding to each child individual in the G child population by using the iterative objective function;
8.6, comparing the predicted value of the injection and production system corresponding to each child individual in the G child population with the weighted average net present value corresponding to the injection and production system corresponding to the initial population one by one according to the child population numbering sequence, and when the predicted value of the injection and production system corresponding to the child individual in the child population is larger than the weighted average net present value of the injection and production system corresponding to the individual in the initial population, replacing the individual in the initial population by the child individual in the child population to update the initial population;
and 8.7, determining the optimal individuals in the updated initial population, acquiring an optimal injection and production system Xb and a weighted average net present value corresponding to the optimal individuals, updating the optimal injection and production system Xb and iteration times, and returning to the step 7.
6. The machine learning offline proxy model production optimization method for reducing development risk according to claim 5, wherein in step 9, the method specifically comprises the following steps:
Step 9.1, respectively calculating the optimal individuals in the initial population and the databaseThe Euclidean distance between the sample data in the database is increased from small to large>The sample data in the sequence are ordered, the first A sample data in the sequence are selected, and a kriging proxy model is constructed>As shown in formula (8):
（8）
in the method, in the process of the invention,for the kriging proxy model->Degree of contrast and harvest->Predicted value of +.>For the average constant of the sample data, +.>Is an autocorrelation error term with zero mean;
step 9.2, constructing a population Plocal which is identical to the initial population, and carrying out differential evolution treatment on the population Plocal to generate a offspring population，/>Wherein->For offspring population->The%>A child generation individual;
the differential evolution processing process is as follows:
（9）
（10）
in the method, in the process of the invention,for offspring population->Middle->The variation vector of the injection and production system corresponding to each individual, < >>For offspring population->Middle->The +.about.th of the variation vector corresponding to each individual>Dimension; />Is a variation factor with a value range of [0.4,1 ]]；/>、/>、/>The injection and production rates corresponding to three mutually different individuals in the population Plocal; />For offspring population->Middle->The +.f. of the injection and production system corresponding to each offspring individual>Dimension; />Is the >First digit of random vector>Dimension; />Is constant, between 0 and 1; />An index selected randomly;
step 9.3, performing differential evolution processing on the population Plocal for a plurality of times according to the preset differential evolution processing times, wherein in each differential evolution processing process, a kriging proxy model is utilizedCalculating offspring population->Predictive value of injection and production system corresponding to each offspring individual, current offspring population +.>When the predicted value of the injection and production system corresponding to the child individuals in the population Plocal is smaller than the weighted average net present value of the injection and production system corresponding to the corresponding individuals in the population Plocal, the child population +.>The offspring individuals in the population Plocal are replaced by individuals in the population Plocal, and the species is updatedGroup Plocal;
step 9.4, searching the population Plocal after the differential evolution treatment by using an iterative objective function, and determining the optimal individual in the population Plocal after the differential evolution treatmentCalculating optimal individuals using iterative objective functions>The predicted value of the corresponding injection and production system is weighted average net present value of the optimal injection and production degree Xb corresponding to the optimal individual in the initial population>Comparing; if the optimal individuals in the population Plocal +.>The predicted value of the corresponding injection and production system is smaller than the weighted average net present value of the optimal injection and production system Xb corresponding to the optimal individual in the initial population >Then use the optimal individuals in the population Plocal +.>And (3) replacing the optimal individuals in the initial population, returning to the step (7) after updating the initial population and the iteration times, otherwise, directly updating the iteration times, and returning to the step (7).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311490309.9A CN117236195B (en) | 2023-11-10 | 2023-11-10 | Machine learning offline agent model production optimization method for reducing development risk |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311490309.9A CN117236195B (en) | 2023-11-10 | 2023-11-10 | Machine learning offline agent model production optimization method for reducing development risk |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117236195A CN117236195A (en) | 2023-12-15 |
| CN117236195B true CN117236195B (en) | 2024-02-02 |
Family
ID=89082858
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311490309.9A Active CN117236195B (en) | 2023-11-10 | 2023-11-10 | Machine learning offline agent model production optimization method for reducing development risk |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117236195B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145278A (en) * | 2016-03-01 | 2017-09-08 | 阿里巴巴集团控股有限公司 | Apparatus control method, device and mobile terminal based on mobile terminal |
| CN108781235A (en) * | 2017-06-13 | 2018-11-09 | 华为技术有限公司 | A kind of display methods and device |
| CN111625922A (en) * | 2020-04-15 | 2020-09-04 | 中国石油大学(华东) | Large-scale oil reservoir injection-production optimization method based on machine learning agent model |
| CN111861774A (en) * | 2020-06-22 | 2020-10-30 | 中国石油大学(华东) | Oil reservoir production machine learning method based on parallel agent model |
| CN111861129A (en) * | 2020-06-22 | 2020-10-30 | 中国石油大学(华东) | Multi-fidelity injection-production optimization method based on multi-scale oil reservoir model |
| CN116205164A (en) * | 2023-04-27 | 2023-06-02 | 中国石油大学(华东) | Multi-agent injection and production optimization method based on self-adaptive basis function selection |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114896903A (en) * | 2022-05-07 | 2022-08-12 | 北京科技大学 | Forced learning-based decision optimization method for oil field production system |
-
2023
- 2023-11-10 CN CN202311490309.9A patent/CN117236195B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145278A (en) * | 2016-03-01 | 2017-09-08 | 阿里巴巴集团控股有限公司 | Apparatus control method, device and mobile terminal based on mobile terminal |
| CN108781235A (en) * | 2017-06-13 | 2018-11-09 | 华为技术有限公司 | A kind of display methods and device |
| CN111625922A (en) * | 2020-04-15 | 2020-09-04 | 中国石油大学(华东) | Large-scale oil reservoir injection-production optimization method based on machine learning agent model |
| CN111861774A (en) * | 2020-06-22 | 2020-10-30 | 中国石油大学(华东) | Oil reservoir production machine learning method based on parallel agent model |
| CN111861129A (en) * | 2020-06-22 | 2020-10-30 | 中国石油大学(华东) | Multi-fidelity injection-production optimization method based on multi-scale oil reservoir model |
| CN116205164A (en) * | 2023-04-27 | 2023-06-02 | 中国石油大学(华东) | Multi-agent injection and production optimization method based on self-adaptive basis function selection |
Non-Patent Citations (2)
| Title |
|---|
| Zhao M, Zhang K, Chen G等.A classification-based surrogateassisted multiobjective evolutionary algorithm for production optimization under geological uncertainty.《SPE Journal,》.2020,2450-2469. * |
| 基于主成分分析和代理模型的油藏生产注采优化方法;张凯;陈国栋;薛小明;张黎明;孙海;姚传进;;中国石油大学学报(自然科学版)(第03期);10042-10049 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117236195A (en) | 2023-12-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111898689B (en) | Image classification method based on neural network architecture search | |
| CN109472321B (en) | Time series type earth surface water quality big data oriented prediction and evaluation model construction method | |
| CN115906675B (en) | Well position and injection and production parameter joint optimization method based on time sequence multi-target prediction model | |
| CN107516135A (en) | A kind of automation monitoring learning method for supporting multi-source data | |
| CN109685653A (en) | A method of fusion deepness belief network and the monitoring of the credit risk of isolated forest algorithm | |
| CN116205164B (en) | Multi-agent injection and production optimization method based on self-adaptive basis function selection | |
| CN109165672A (en) | A kind of Ensemble classifier method based on incremental learning | |
| CN110807544A (en) | Oil field residual oil saturation distribution prediction method based on machine learning | |
| CN114528949A (en) | Parameter optimization-based electric energy metering abnormal data identification and compensation method | |
| CN104732067A (en) | Industrial process modeling forecasting method oriented at flow object | |
| CN116542429A (en) | Oil reservoir production index machine learning prediction method integrating space-time characteristics | |
| CN113515891B (en) | Emulsion explosive quality prediction and optimization method | |
| CN116882323B (en) | Self-adaptive agent policy optimization method considering time sequence and subdivision tasks | |
| CN116680548B (en) | Time sequence drought causal analysis method for multi-source observation data | |
| CN117236195B (en) | Machine learning offline agent model production optimization method for reducing development risk | |
| CN115618987A (en) | Production well production data prediction method, device, equipment and storage medium | |
| CN109409434A (en) | The method of liver diseases data classification Rule Extraction based on random forest | |
| CN113095501A (en) | Deep reinforcement learning-based unbalanced classification decision tree generation method | |
| CN117334271A (en) | Method for generating molecules based on specified attributes | |
| CN112214524A (en) | Data evaluation system and evaluation method based on deep data mining | |
| CN116303626A (en) | Well cementation pump pressure prediction method based on feature optimization and online learning | |
| CN116720743A (en) | Carbon emission measuring and calculating method based on data clustering and machine learning | |
| CN113837096B (en) | Rolling bearing fault diagnosis method based on GA random forest | |
| CN114862007A (en) | Short-period gas production rate prediction method and system for carbonate gas well | |
| CN114004170A (en) | Reconstruction method of cascade flow field based on limited measuring points |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |