A production optimization method for porous, thick carbonate reservoirs

Abstract

This application discloses a production optimization method for porous, thick carbonate reservoirs, relating to the field of petroleum engineering technology. First, a spatiotemporal inference surrogate set is constructed and coupled with a mass conservation loss function based on gravity terms. The spatiotemporal inference surrogate set is then trained using sample training data and the mass conservation loss function. Next, the trained spatiotemporal inference surrogate set is used to analyze reservoir production candidate schemes, generating predicted reservoir production index data. Based on the predicted reservoir production index data, the sample training data undergoes inner-loop zero-cost sample expansion and outer-loop high-cost numerical simulation calibration. Finally, combining economic parameters and the cumulative oil production objective function value, the reservoir production candidate schemes are updated using non-dominated ranking, crowding distance screening, and genetic operations. This process is repeated iteratively until the termination condition is met, outputting the target reservoir production scheme. This method solves the problems of high cost, black-box defects, and sample scarcity inherent in traditional techniques, improving the efficiency and reliability of reservoir production optimization.

Description

Technical Field

[0001] This application relates to the field of petroleum engineering technology, specifically to a method for optimizing the production of porous, thick carbonate reservoirs. Background Technology

[0002] In the waterflooding development lifecycle of porous, thick carbonate reservoirs, replenishing formation energy through water injection is the core means of enhancing oil recovery. Due to the enormous reservoir thickness (usually exceeding 100 meters), there is extremely strong vertical / lateral heterogeneity and gravity differentiation effect in the matrix pores. To maximize net present value (NPV) and cumulative production (COP), traditional closed-loop reservoir management typically relies on fully implicit finite difference numerical simulators for multi-objective production optimization (MOPO), which constitutes a typical high-dimensional, strongly nonlinear, multi-objective engineering optimization problem.

[0003] Traditional full-physics numerical simulation faces insurmountable computational bottlenecks. Traditional closed-loop reservoir production optimization heavily relies on full-physics numerical simulators, which solve partial differential equations for the seepage of multiphase fluids in porous media using finite difference or finite volume methods. While these simulators offer high computational accuracy, the mesh size of detailed three-dimensional geological models of porous, thick carbonate rocks typically reaches millions or even tens of millions of pixels. In multi-objective production optimization, evolutionary algorithms (such as genetic algorithms and particle swarm optimization) often require tens of thousands of calls to the numerical simulator to evaluate the fitness of candidate injection-production schemes. A single full-scale model simulation can easily take hours or even days. This extremely high computational cost makes real-time, high-dimensional injection-production optimization completely impractical in real-world engineering.

[0004] The physical inconsistencies and black-box prediction oscillations inherent in purely data-driven spatiotemporal inference agents present significant challenges. To overcome the aforementioned computational bottlenecks, the industry has introduced spatiotemporal inference agent technology in recent years, utilizing machine learning algorithms (such as artificial neural networks, support vector regression, and radial basis functions) to replace time-consuming numerical simulators. While purely data-driven spatiotemporal inference agents have greatly improved evaluation efficiency, they essentially establish a purely mathematical mapping between inputs (well control parameters) and outputs (production capacity or physical fields), resulting in a fatal black-box defect. When dealing with porous, thick reservoirs, the vertical gravity differentiation effect of fluids plays a decisive role in the advancement of the water drive front. Without physical constraints, purely data-driven models are prone to producing predictions that violate Darcy's law of flow and the principle of mass conservation. Particularly when predicting three-dimensional pressure fields and water saturation fields, non-physical numerical oscillations and severe mass balance errors frequently occur, preventing reservoir engineers from developing safe and reliable injection-production control strategies.

[0005] The dimensionality curse and sample scarcity paradox in high-dimensional decision spaces: Production optimization in porous, thick oil reservoirs often involves parameter decisions for dozens of wells over multiple control steps spanning several years, with extremely high-dimensional decision variables (often reaching hundreds of dimensions). Training a high-precision spatiotemporal inference agent in such a high-dimensional space requires an exponentially increasing number of high-fidelity training samples. This leads to an engineering paradox: if relying entirely on high-fidelity numerical simulators to obtain a large number of samples, the computational cost remains astronomical, defeating the purpose of cost reduction and efficiency improvement for the spatiotemporal inference agent; while relying only on a small number of initial samples (such as traditional Latin hypercube sampling), a single spatiotemporal inference agent will produce severe inductive bias, resulting in distorted predictions when faced with unknown parameter combinations. Summary of the Invention

[0006] The purpose of this application is to provide a production optimization method for porous, thick carbonate reservoirs, which solves the problems of high cost, black box defects and / or difficulty in obtaining massive samples in existing technologies.

[0007] This application is achieved through the following technical solution:

[0008] A method for optimizing production in porous, thick carbonate reservoirs includes:

[0009] Sample geological well control parameters and corresponding ground value label data of the sample geological well control parameters are collected during the production process of porous, thick carbonate reservoirs to obtain sample training data; the ground value label data represents reservoir production index data.

[0010] Construct a set of spatiotemporal inference agents and a matter conservation loss function coupled with a gravity term; the set of spatiotemporal inference agents includes at least one spatiotemporal inference agent;

[0011] Based on the sample training data and the matter conservation loss function, the spatiotemporal inference agent set is trained to obtain the trained spatiotemporal inference agent set.

[0012] Generate reservoir production candidate schemes, and use the trained spatiotemporal inference agent set to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data.

[0013] Based on the predicted reservoir formation index data, the sample training data is expanded by using inner-loop zero-cost sample expansion based on Tri-training collaborative training and outer-loop high-cost numerical simulation calibration based on active learning agent set query, and the expanded sample training data is obtained.

[0014] Based on the predicted reservoir production index data, the objective function values ​​of economic parameters and cumulative oil production are obtained. Then, based on the objective function values ​​of economic parameters and cumulative oil production, non-dominated sorting, crowding distance screening, and genetic operations are performed on the reservoir production candidate schemes to obtain updated reservoir production candidate schemes.

[0015] The process involves iterating repeatedly using expanded and calibrated sample training data and updated reservoir production candidate schemes until the iteration termination condition is met. Then, based on the reservoir production candidate schemes after non-dominated sorting in the last iteration, the target reservoir production scheme for porous, thick carbonate reservoirs is determined.

[0016] In one possible implementation, the sample geological well control parameters include a three-dimensional static geological attribute field, effective reservoir thickness, the ratio of vertical to horizontal permeability characterizing vertical heterogeneity, and a vector of well control decision variables.

[0017] In one possible implementation, a spatiotemporal inference agent set is constructed, including:

[0018] A first, second, and third spatiotemporal inference agent, all of which are CNN-LSTM models, are constructed to obtain a spatiotemporal inference agent set. The output of the spatiotemporal inference agent set is obtained by weighted summation of the outputs of the first, second, and third spatiotemporal inference agents.

[0019] In one possible implementation, the mass conservation loss function coupled with the gravity term is constructed as follows:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] in, The mass conservation loss function representing the coupled gravity term. This represents the mean square error term. Indicates the weight of the data item. Indicates the physical constraint penalty weight. This indicates the number of spatiotemporal coordinate points within the computational domain. This represents the partial differential parameter of the oil phase at the z-th spatiotemporal collocation point. This represents the partial differential parameter of the oil phase at the z-th spatiotemporal collocation point. The function representing the calculation of the partial differential parameters of the aqueous phase. The function representing the calculation of the partial differential parameters of the oil phase. Represents the differential symbol. Indicates porosity. This represents the density of the aqueous phase. Indicates time, Represents the gradient. Indicates the water phase flow rate. Indicates oil phase flow rate, Represents the velocity vector of water phase seepage. Indicates the density of the oil phase. This represents the oil phase flow velocity vector. The function representing the calculation of the velocity vector of water phase seepage. The function representing the calculation of the oil phase seepage velocity vector. Indicates spatial location The absolute permeability tensor at that location, This indicates the spatial location of spatiotemporal points or grid centers within the three-dimensional computational domain of porous, thick carbonate reservoirs. Indicates the relative permeability of water. Indicates the viscosity of the aqueous phase. Indicates spatial location r and time The corresponding predicted pressure field value, Represents spatial coordinate vectors The gradient operator, where g represents gravitational acceleration. Spatial gradient in the direction of gravity Indicates the relative permeability of oil. Indicates the viscosity of the oil phase. This represents the predicted value of the oil saturation field. This represents the predicted value of the water saturation field.

[0027] In one possible implementation, the spatiotemporal inference agent set is trained based on the sample training data and the matter conservation loss function to obtain the trained spatiotemporal inference agent set, including:

[0028] Based on the sample training data and with the goal of minimizing the matter conservation loss function, the spatiotemporal inference agent set is trained using the gradient descent algorithm to obtain the trained spatiotemporal inference agent set.

[0029] In one possible implementation, reservoir production candidate schemes are generated, and the trained spatiotemporal inference proxy set is used to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data, including:

[0030] Generate reservoir production candidate schemes; the reservoir production candidate schemes include a vector of candidate well control decision variables;

[0031] The three-dimensional static geological property field, effective reservoir thickness, and vertical to horizontal permeability ratio characterizing vertical heterogeneity were collected during the production process of porous, thick carbonate reservoirs, and these were combined with candidate reservoir production schemes to form real-time geological well control parameters.

[0032] The real-time geological well control parameters are used as input to the trained spatiotemporal inference agent set, and the predicted reservoir production index data output by the trained spatiotemporal inference agent set are obtained.

[0033] The predicted reservoir production index data includes the predicted reservoir production index output by the first spatiotemporal inference agent, the second spatiotemporal inference agent, the third spatiotemporal inference agent, and the spatiotemporal inference agent set.

[0034] In one possible implementation, based on the predicted reservoir formation index data, the sample training data is expanded using an inner-loop zero-cost sample expansion based on Tri-training collaborative training and an outer-loop high-cost numerical simulation calibration based on active learning agent set query, to obtain the expanded sample training data, including:

[0035] Based on the predicted reservoir generation index data, the maximum divergence of the reservoir production candidate schemes is evaluated, and the maximum divergence of the reservoir production candidate schemes is obtained as follows:

[0036] ;

[0037] in, This represents the maximum degree of divergence corresponding to the expanded sample data. This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... Predicted reservoir production indicators obtained by a spatiotemporal inference agent; This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... The predicted reservoir production indicators obtained by the spatiotemporal inference agent; max represents the maximum value function. =1,2,3; =1,2,3;

[0038] A high-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is less than a pre-set divergence threshold.

[0039] A low-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is greater than or equal to a pre-set divergence threshold.

[0040] Based on the high-confidence candidate sample set, perform zero-cost inner-loop sample expansion using Tri-training collaborative training, and perform high-cost outer-loop numerical simulation calibration based on active learning agent set query using the low-confidence candidate sample set to obtain expanded sample data;

[0041] The original sample training data is expanded using the expanded sample data to obtain the expanded sample training data.

[0042] In one possible implementation, inner-loop zero-cost sample augmentation based on Tri-training co-training is performed according to the high-confidence candidate sample set, including:

[0043] For any high-confidence reservoir production candidate scheme in the high-confidence candidate sample set, obtain the mean value of the predicted reservoir production index corresponding to the predicted reservoir production index output by the two spatiotemporal inference agents that participated in the calculation of the maximum divergence degree;

[0044] Based on the high-confidence reservoir production candidate schemes, sample geological well control parameters are constructed, and the average predicted reservoir production indexes are used to construct corresponding pseudo-label data to obtain the first expanded sample data.

[0045] In one possible implementation, high-cost numerical simulation calibration of the outer loop based on an active learning agent set query is performed according to a set of low-confidence candidate samples, including:

[0046] For any low-confidence reservoir production candidate scheme in the set of low-confidence candidate samples, obtain the prediction physics field (PDE) residual corresponding to the low-confidence reservoir production candidate scheme. Pareto Front Proximity Indicator Function and the decision space distance with existing high-fidelity samples. ; This represents a low-confidence reservoir production candidate scheme in the low-confidence candidate sample set; the existing high-fidelity sample refers to the existing second augmented sample data;

[0047] The maximum degree of divergence corresponding to the production candidate schemes for the low-confidence reservoir. Predicting the PDE residuals of physical fields and decision space distance Normalization was performed separately to obtain the maximum divergence degree of the normalized surrogate model. Normalized prediction physical field PDE residuals and normalized decision space distance ;

[0048] Based on the maximum divergence of the normalized agency model Normalized prediction physical field PDE residuals Normalized decision space distance and the Pareto front proximity indicator function Calculate the production candidate schemes for the low-confidence reservoir. The high-fidelity calibration value function value is:

[0049] ;

[0050] in, Indicates low-confidence reservoir production candidate schemes High-fidelity calibration value function value. , , , Let these represent the first weight, second weight, third weight, and fourth weight, respectively, and satisfy the following conditions: , ;

[0051] According to the high-fidelity calibration value function value The low-confidence reservoir production candidate schemes in the low-confidence candidate sample set are sorted from largest to smallest, and the schemes ranked first are selected. Low-confidence reservoir production candidate schemes are used as calibration samples; or schemes that meet the following criteria are selected. Low-confidence reservoir production candidate schemes were used as calibration samples; among them, This indicates the maximum number of samples allowed to be called from the reservoir numerical simulator in each iteration. This indicates a preset high-fidelity calibration value threshold;

[0052] The reservoir numerical simulator is called to perform simulation calibration on the sample to be calibrated to obtain the corresponding real reservoir production index data.

[0053] The geological well control parameters of the sample to be calibrated are constructed, and the real reservoir production index data corresponding to the sample to be calibrated are used as the corresponding high-fidelity label data to obtain the second expanded sample data.

[0054] In one possible implementation, based on the predicted reservoir production index data, the objective function values ​​of economic parameters and cumulative oil production are obtained. Then, based on these objective function values, the reservoir production candidate schemes are subjected to non-dominated ranking, crowding distance screening, and genetic operations to obtain updated reservoir production candidate schemes, including:

[0055] Based on the predicted reservoir production index data, the objective function values ​​for economic parameters and cumulative oil production are obtained as follows:

[0056] ;

[0057] ;

[0058] in, This represents the objective function value of the economic parameters. This represents the objective function value of cumulative oil production. Indicates total production time steps; This represents the length of the nth time step. Indicates the discount rate. Indicates oil price, Indicates the price of water treatment. Indicates the total number of producing wells. Indicates the total number of injection wells. This represents the oil production of the j-th producing well in the nth time period. This represents the water production of the j-th production well in the nth time period. Indicates the price of water injection. Let i represent the water injection volume of the j-th injection well in the n-th time period, and i represent the injection well.

[0059] Based on the objective function values ​​of the economic parameters and the objective function value of the cumulative oil production, the reservoir production candidate schemes are subjected to non-dominated sorting, crowding distance screening, and genetic operations to obtain updated reservoir production candidate schemes.

[0060] Compared with the prior art, this application has the following advantages and beneficial effects:

[0061] This application provides a production optimization method for porous, thick carbonate reservoirs. First, a spatiotemporal inference surrogate set is constructed and coupled with a mass conservation loss function based on gravity terms. The spatiotemporal inference surrogate set is then trained using sample training data and the mass conservation loss function. Next, the trained spatiotemporal inference surrogate set is used to analyze reservoir production candidate schemes, generating predicted reservoir production index data. Based on the predicted reservoir production index data, the sample training data undergoes inner-loop zero-cost sample expansion and outer-loop high-cost numerical simulation calibration. Finally, combining economic parameters and the cumulative oil production objective function value, the reservoir production candidate schemes are updated using non-dominated ranking, crowding distance screening, and genetic operations. This process is repeated iteratively until the termination condition is met, outputting the target reservoir production scheme. This method solves the problems of high cost, black-box defects, and sample scarcity inherent in traditional techniques, improving the efficiency and reliability of reservoir production optimization. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the exemplary embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0063] Figure 1 A flowchart illustrating a production optimization method for porous, thick carbonate reservoirs provided in this application embodiment.

[0064] Figure 2 A schematic diagram of a full-size reservoir numerical simulation model provided in the embodiments of this application;

[0065] Figure 3 A schematic diagram of the final output Pareto optimization solution set provided in the embodiments of this application;

[0066] Figure 4 A schematic diagram of the production well bottom pressure combination provided in an embodiment of this application; wherein, Figure 4 (a) in the diagram is a schematic diagram of the bottom pressure of the production well in Scheme 1. Figure 4 (b) in the diagram is a schematic diagram of the bottom pressure of the production well in Scheme 2;

[0067] Figure 5 This is a schematic diagram of the injection volume combination of the injection well provided in an embodiment of this application; wherein, Figure 5 (a) in the diagram is a schematic diagram of the injection volume of the injection well in Scheme 1. Figure 5 (b) in the diagram is a schematic diagram of the injection volume of the injection well in Scheme 2;

[0068] Figure 6A schematic diagram comparing NPV and COP between Scheme 1 and Scheme 2 provided in the embodiments of this application. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0070] Due to the high cost, black-box defects, and / or difficulty in obtaining massive samples in existing technologies, current reservoir production optimization techniques cannot achieve a balance between extremely low computational budgets, high-dimensional parameter optimization, and strict adherence to fluid physics laws. To address these issues, this application provides a dual-drive spatiotemporal inference agent embedded with the gravity physics mechanism of oil-water two-phase systems. It constructs a dual-loop management architecture with zero-cost expansion of the inner loop pseudo-labels and minimal active calibration of the outer loop truth values. While maintaining millisecond-level supercomputing efficiency, it completely solves the physical inconsistency problem of the spatiotemporal inference agent, enabling high-dimensional, high-precision multi-objective injection-production dynamic optimization of porous, thick carbonate reservoirs under extremely limited computing power.

[0071] like Figure 1 As shown in the embodiments of this application, a method for optimizing the production of porous, thick carbonate reservoirs is provided, including:

[0072] S101. Collect sample geological well control parameters and the corresponding true value label data of the sample geological well control parameters during the production process of porous, thick carbonate reservoirs to obtain sample training data; the true value label data represents reservoir production index data.

[0073] S102. Construct a set of spatiotemporal inference agents and a matter conservation loss function coupled with the gravity term; the set of spatiotemporal inference agents includes at least one spatiotemporal inference agent;

[0074] S103. Based on the sample training data and the matter conservation loss function, train the spatiotemporal inference agent set to obtain the trained spatiotemporal inference agent set.

[0075] S104. Generate reservoir production candidate schemes, and use the trained spatiotemporal inference agent set to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data.

[0076] S105. Based on the predicted reservoir generation index data, the sample training data is expanded by using inner-loop zero-cost sample expansion based on Tri-training collaborative training and outer-loop high-cost numerical simulation calibration based on active learning agent set query, and the expanded sample training data is obtained.

[0077] S106. Based on the predicted reservoir production index data, obtain the objective function value of the economic parameter and the objective function value of the cumulative oil production, and perform non-dominated sorting, crowding distance screening and genetic operation on the reservoir production candidate scheme according to the objective function value of the economic parameter and the objective function value of the cumulative oil production to obtain the updated reservoir production candidate scheme.

[0078] S107. The expanded and calibrated sample training data and the updated reservoir production candidate schemes are used for repeated iterations until the iteration termination condition is met. Then, the target reservoir production scheme for the production of porous, thick carbonate reservoirs is determined based on the reservoir production candidate schemes after non-dominated sorting in the last iteration.

[0079] Optionally, the iteration termination condition includes at least one of the following: the current iteration number reaches the preset maximum number of iterations; the cumulative number of high-fidelity numerical simulation calls reaches the preset upper limit; the change rate of the Pareto front hypervolume index is less than a preset threshold for several consecutive generations; the improvement of the normalized objective function value is less than a preset threshold for several consecutive generations; and the maximum divergence of the surrogate model and the PDE residual of the predicted physics field in the current reservoir production candidate scheme set are both less than the corresponding preset thresholds.

[0080] In one possible implementation, the sample geological well control parameters include a three-dimensional static geological attribute field, effective reservoir thickness, the ratio of vertical to horizontal permeability characterizing vertical heterogeneity, and a vector of well control decision variables.

[0081] For example, sample geological well control parameters and corresponding ground truth label data pre-stored from a specified data source can be obtained as sample training data. Alternatively, sample geological well control parameters can be collected first, and then a reservoir numerical simulator can be used for simulation to output the actual development response, which can then be used as the ground truth label data.

[0082] To address the strong vertical and horizontal heterogeneity within porous, thick carbonate reservoirs, multi-scale static physical parameters can be extracted from fine three-dimensional geological models with millions or even tens of millions of grids. The core geological feature matrix extracted includes a three-dimensional static geological attribute field (porosity ϕ, three-dimensional absolute permeability tensor K).

[0083] Considering that fluids in porous, thick carbonate reservoirs are prone to phenomena such as bottom water inrush, in order for the subsequent physical constraint neural network to accurately capture the gravity differentiation effect, it is necessary to perform depth feature extraction and grid coarsening mapping on the effective reservoir thickness and the ratio of vertical to horizontal permeability (Kv / Kh) that characterizes vertical heterogeneity, and convert them into static background tensors.

[0084] For an oil reservoir containing P production wells and I water injection wells, the entire lifecycle production process is discretized into N. t For each control time step n, define the well control decision variable vector u. (n) as follows:

[0085] ;

[0086] in, Indicates the first The injection well at the first The amount of water injected at each control time step. ; Indicates the first The first oil well in the country Bottom-hole flowing pressure at each control time step ; superscript This indicates the transpose of a matrix or vector.

[0087] Water injection volume of injection well Used to control formation energy replenishment at the injection end and water drive front propulsion; bottomhole flowing pressure in oil wells. Used to control the pressure differential and fluid production intensity at the production end. Both serve as dynamic well control variables input to the spatiotemporal inference surrogate model, and are combined with the static geological tensor through multi-channel splicing or feature embedding to form the input feature space of the deep learning model. Based on the above inputs, the spatiotemporal inference surrogate model predicts the pressure saturation field, water saturation field, oil saturation field, oil production, and water production of each well at the corresponding control time step, and then uses this prediction to calculate the net present value and cumulative oil production objective functions.

[0088] In one possible implementation, a spatiotemporal inference agent set is constructed, including:

[0089] A first, second, and third spatiotemporal inference agent, all of which are CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) models, are constructed to obtain a spatiotemporal inference agent set. The output of the spatiotemporal inference agent set is obtained by weighted summation of the outputs of the first, second, and third spatiotemporal inference agents.

[0090] To overcome the shortcomings of single data-driven models in high-dimensional injection-mapping decision spaces, which are prone to local bias and excessive variance, this application introduces the Bagging concept from ensemble learning. The aim is to minimize the variance of model predictions by constructing a parallel set of multi-model surrogates, while laying the foundation for subsequent active learning and collaborative training.

[0091] The Bootstrap resampling technique was used to randomly divide the limited initial high-fidelity historical dataset into three distinct sub-training sets. An architecture based on a coupling of convolutional and long short-term memory networks (CNN-LSTM) was independently built for each sub-set, forming a spatiotemporal inference proxy set containing three spatiotemporal inference models.

[0092] To address the dynamic inversion of the three-dimensional spatial field, a convolutional recurrent neural network (CRNN) is constructed. At the control time step n, a three-dimensional convolutional layer (3D-CNN) is used to extract local features of the longitudinal and transverse heterogeneous space of the thick oil reservoir. These features are then input into a long short-term memory network (LSTM) to perform Markov time-series evolution inference of fluid flow.

[0093] The hidden state update mechanism of LSTM cells is characterized as follows:

[0094] ;

[0095] in, Indicates the first The LSTM hidden states corresponding to the nth control time step are used to characterize the state up to the nth control time step. Time-series characteristics of reservoir dynamic response at each control time step; Indicates the first The LSTM cell state corresponding to each control time step is used to store and transmit fluid transport and well control response memory information over long time scales; The static geological characteristics tensor representing porous, thick carbonate reservoirs includes static geological attributes such as porosity field, permeability field, effective reservoir thickness, and the ratio of vertical to horizontal permeability. Indicates the first The well control decision variable vector under each control time step includes dynamic control parameters such as the injection volume of each injection well and the bottom flow pressure of each production well; Indicates the first The LSTM hidden state corresponding to each control time step; Indicates the first The LSTM cell state corresponding to each control time step.

[0096] Using the three independently trained spatiotemporal inference agents mentioned above, a spatiotemporal inference agent set is constructed. When predicting the production capacity or physical field of unknown injection and extraction schemes, the final output response fens(x) of the agent set is not a simple arithmetic average, but rather adopts an adaptive weighted summation mechanism based on the current root mean square error (RMSE) of each model:

[0097] ;

[0098] ;

[0099] in, This represents the output of the spatiotemporal inference agent set. This represents the output of the first spatiotemporal inference agent. This represents the output of the second spatiotemporal inference agent. This represents the output of the third spatiotemporal inference agent. This represents the dynamic aggregation weight of the first spatiotemporal inference agent. This represents the dynamic aggregation weights of the second spatiotemporal inference agent. This represents the dynamic aggregation weight of the third spatiotemporal inference agent. Indicates the first The root mean square error (RMSE) of the spatiotemporal inference agents on the validation set, which refers to a fixed number of data points randomly selected from the sample training data, is used. This adaptive weighting mechanism ensures that spatiotemporal inference agents with higher local accuracy have a larger weight in the agent set when predicting the fluid front movement of thick oil reservoirs, thereby significantly improving the robustness of spatiotemporal agent predictions.

[0100] In one possible implementation, after constructing the spatiotemporal inference agent set, in order to break the non-physical prediction oscillations (such as fluid backflow or saturation anomalies that violate Darcy's law) generated by the pure data-driven model when facing complex porous and thick carbonate rocks, the embodiments of this application activate the automatic differentiation mechanism at the bottom layer of the deep learning framework and specifically extract the hydrodynamic partial differential (PDE) residuals coupled with gravitational potential energy.

[0101] Utilizing the automatic differentiation technique at the bottom layer of deep learning frameworks, spatial coordinate vectors are directly derived from the network output layer. and time Apply chain rule differentiation. Where, This indicates the spatial location of spatiotemporal points or grid centers within the three-dimensional computational domain of porous, thick carbonate reservoirs. , , These represent the coordinate components of the reservoir model in three orthogonal directions; time... This represents a continuous-time variable or the corresponding control time step. Taking aqueous seepage as an example, considering the difference in gravitational potential energy in a thick carbonate rock, the velocity vectors of aqueous and oil phases containing the gravity term are calculated using Darcy's law:

[0102]

[0103]

[0104] Furthermore, the saturation of both phases satisfies the auxiliary conservation constraint:

[0105] ;

[0106] in, Represents spatial coordinate vectors The gradient operator; Indicates spatial location and time Corresponding pressure field prediction values; and These represent the predicted values ​​for water saturation and oil saturation, respectively. Indicates spatial location The corresponding vertical depth or elevation function, Represents the spatial gradient along the direction of gravity; Indicates spatial location The absolute permeability tensor at that location; and These represent the relative permeability functions of the aqueous and oil phases, respectively. and These represent the viscosity of the aqueous phase and the viscosity of the oil phase, respectively. and These represent the densities of the aqueous phase and the oil phase, respectively. It represents the acceleration due to gravity.

[0107] The mass conservation law for both oil and water phases is transformed into a physical constraint penalty term in a neural network. Combined with the extracted flow velocity vector, the mass balance residuals of the partial differential equations (PDEs) for the water and oil phases are defined as follows:

[0108] ;

[0109] ;

[0110] Based on the above parameters, the mass conservation loss function for the coupled gravity term is constructed as follows:

[0111] ;

[0112] in, The mass conservation loss function representing the coupled gravity term. This represents the mean square error term. Indicates the weight of the data item. Indicates the physical constraint penalty weight. This indicates the number of spatiotemporal coordinate points within the computational domain. Denotes the partial differential parameter of the water phase at the z-th spatiotemporal collocation point. This represents the partial differential parameter of the oil phase at the z-th spatiotemporal collocation point. The function representing the calculation of the partial differential parameters of the aqueous phase. The function representing the calculation of the partial differential parameters of the oil phase. Represents the differential symbol. Indicates porosity. This represents the density of the aqueous phase. Indicates time, Represents the gradient. Indicates the water phase flow rate. Indicates oil phase flow rate, Represents the velocity vector of water phase seepage. Indicates the density of the oil phase. Let K represent the oil phase flow velocity vector, and K represent the absolute permeability tensor. Indicates the relative permeability of water. Indicates the viscosity of the aqueous phase. The pressure gradient is represented by g, and g represents the acceleration due to gravity. Indicates the longitudinal position gradient. Indicates the relative permeability of oil. This indicates the viscosity of the oil phase.

[0113] During training, a gradient descent algorithm is typically used based on a matter conservation loss function. Therefore, when the spatiotemporal inference surrogate set violates the fluid transport and matter conservation laws of any phase, Res w or Res o This will generate a huge penalty gradient that is fed back to the network neurons.

[0114] In one possible implementation, the spatiotemporal inference agent set is trained based on the sample training data and the matter conservation loss function to obtain the trained spatiotemporal inference agent set, including:

[0115] Based on the sample training data and with the goal of minimizing the matter conservation loss function, the spatiotemporal inference agent set is trained using the gradient descent algorithm to obtain the trained spatiotemporal inference agent set.

[0116] In one possible implementation, reservoir production candidate schemes are generated, and the trained spatiotemporal inference proxy set is used to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data, including:

[0117] Generate reservoir production candidate schemes; the reservoir production candidate schemes include a vector of candidate well control decision variables;

[0118] The three-dimensional static geological property field, effective reservoir thickness, and vertical to horizontal permeability ratio characterizing vertical heterogeneity were collected during the production process of porous, thick carbonate reservoirs, and these were combined with candidate reservoir production schemes to form real-time geological well control parameters.

[0119] The real-time geological well control parameters are used as input to the trained spatiotemporal inference agent set, and the predicted reservoir production index data output by the trained spatiotemporal inference agent set are obtained.

[0120] The predicted reservoir production index data includes the predicted reservoir production index output by the first spatiotemporal inference agent, the second spatiotemporal inference agent, the third spatiotemporal inference agent, and the spatiotemporal inference agent set.

[0121] In one possible implementation, based on the predicted reservoir formation index data, the sample training data is expanded using an inner-loop zero-cost sample expansion based on Tri-training collaborative training and an outer-loop high-cost numerical simulation calibration based on active learning agent set query, to obtain the expanded sample training data, including:

[0122] Based on the predicted reservoir generation index data, the maximum divergence of the reservoir production candidate schemes is evaluated, and the maximum divergence of the reservoir production candidate schemes is obtained as follows:

[0123] ;

[0124] in, This represents the maximum degree of divergence corresponding to the expanded sample data. This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... Predicted reservoir production indicators obtained by a spatiotemporal inference agent; This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... The predicted reservoir production indicators obtained by the spatiotemporal inference agent; max represents the maximum value function. =1,2,3; =1,2,3;

[0125] A high-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is less than a pre-set divergence threshold.

[0126] A low-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is greater than or equal to a pre-set divergence threshold.

[0127] Based on the high-confidence candidate sample set, perform zero-cost inner-loop sample expansion using Tri-training collaborative training, and perform high-cost outer-loop numerical simulation calibration based on active learning agent set query using the low-confidence candidate sample set to obtain expanded sample data;

[0128] The original training data is augmented using the augmented sample data to obtain augmented training data. In one possible implementation, zero-cost inner-loop sample augmentation based on Tri-training co-training is performed using the high-confidence candidate sample set, including:

[0129] For any high-confidence reservoir production candidate scheme in the high-confidence candidate sample set, obtain the mean value of the predicted reservoir production index corresponding to the predicted reservoir production index output by the two spatiotemporal inference agents that participated in the calculation of the maximum divergence degree;

[0130] Based on the high-confidence reservoir production candidate schemes, sample geological well control parameters are constructed, and the average predicted reservoir production indexes are used to construct corresponding pseudo-label data to obtain the first expanded sample data.

[0131] For example, after the initial sampling is completed, the candidate well control scheme set is sampled, and the numerical simulation is completed, the full physics numerical simulator is disabled, and the Tri-training semi-supervised learning mechanism is started:

[0132] 1. Allow the three spatiotemporal inference agents in the spatiotemporal inference agent set to independently predict reservoir production candidate schemes.

[0133] 2. Establish high confidence screening criteria: If the prediction results of two spatiotemporal inference agents for a certain group of reservoir production candidate schemes are very similar (i.e., the confidence assessment conditions are met), then the reservoir production candidate scheme is considered to have high confidence.

[0134] 3. The predicted mean of the reservoir production candidate scheme with high confidence is used as pseudo-label data, and its calculation formula is as follows:

[0135] ;

[0136] 4. The augmented sample data with pseudo-labels is directly injected into the sample dataset for model retraining. During this process, the physical residual term will spontaneously perform physical mechanism-level correction on the numerical noise that may be carried by the prediction results, realizing the massive data expansion of the high-dimensional decision space at zero computing cost.

[0137] In one possible implementation, high-cost numerical simulation calibration of the outer loop based on an active learning agent set query is performed according to a set of low-confidence candidate samples, including:

[0138] For any low-confidence reservoir production candidate scheme in the set of low-confidence candidate samples, obtain the prediction physics field (PDE) residual corresponding to the low-confidence reservoir production candidate scheme. Pareto Front Proximity Indicator Function and the decision space distance with existing high-fidelity samples. ; This represents a low-confidence reservoir production candidate scheme in the low-confidence candidate sample set; the existing high-fidelity sample refers to the existing second augmented sample data;

[0139] The maximum degree of divergence corresponding to the production candidate schemes for the low-confidence reservoir. Predicting the PDE residuals of physical fields and decision space distance Normalization was performed separately to obtain the maximum divergence degree of the normalized surrogate model. Normalized prediction physical field PDE residuals and normalized decision space distance ;

[0140] Based on the maximum divergence of the normalized agency model Normalized prediction physical field PDE residuals Normalized decision space distance and the Pareto front proximity indicator function Calculate the production candidate schemes for the low-confidence reservoir. The high-fidelity calibration value function value is:

[0141] ;

[0142] in, Indicates low-confidence reservoir production candidate schemes High-fidelity calibration value function value. , , , Let these represent the first weight, second weight, third weight, and fourth weight, respectively, and satisfy the following conditions: , ;

[0143] According to the high-fidelity calibration value function value The low-confidence reservoir production candidate schemes in the low-confidence candidate sample set are sorted from largest to smallest, and the schemes ranked first are selected. Low-confidence reservoir production candidate schemes are used as calibration samples; or schemes that meet the following criteria are selected. Low-confidence reservoir production candidate schemes were used as calibration samples; among them, This indicates the maximum number of samples allowed to be called from the reservoir numerical simulator in each iteration. This indicates a preset high-fidelity calibration value threshold;

[0144] The reservoir numerical simulator is called to perform simulation calibration on the sample to be calibrated to obtain the corresponding real reservoir production index data.

[0145] The geological well control parameters of the sample to be calibrated are constructed, and the real reservoir production index data corresponding to the sample to be calibrated are used as the corresponding high-fidelity label data to obtain the second expanded sample data.

[0146] Optional, Pareto front proximity indicator function The methods of obtaining it include:

[0147] Production candidate schemes for low-confidence reservoirs The objective function values ​​are normalized to obtain the normalized objective function vector:

[0148] ;

[0149] in, Indicates the number of objective functions. Indicates the first Normalized values ​​of the objective functions;

[0150] Let the solution set of the first non-dominated layer under the current iterative algebra be... Calculate candidate production schemes for reservoirs Minimum Euclidean distance to the solution set of the first non-dominated layer:

[0151] ;

[0152] like Then determine the candidate reservoir production schemes. Located in the area adjacent to the frontier of Pareto, and ordered:

[0153] ;

[0154] like Then determine the candidate reservoir production schemes. Not located in the vicinity of the Pareto frontier, and therefore:

[0155] ;

[0156] in, This is the preset Pareto front proximity threshold.

[0157] The decision space distance The methods of obtaining it include:

[0158] Candidate solutions for reservoir production The decision variables are normalized:

[0159]

[0160] in, Indicates candidate reservoir production schemes The One decision variable, and They represent the first The lower and upper bounds of each decision variable. Represents the normalized i-th One decision variable, To prevent extremely small positive numbers with a denominator of zero;

[0161] Suppose we have a set of high-fidelity samples:

[0162] ;

[0163] in, This indicates the number of high-fidelity samples already available. High-fidelity samples refer to samples that have undergone high-fidelity reservoir numerical simulation calibration and obtained real reservoir production index data.

[0164] Selection of reservoir production candidate schemes The closest in the normalized decision space A set of high-fidelity samples constitutes the nearest neighbor set. And calculate:

[0165] ;

[0166] in, Indicates candidate reservoir production schemes The normalized decision variable vector, This indicates that high-fidelity samples are already available. The normalized decision variable vector, Indicates the dimension of decision variables. This indicates the preset number of nearest neighbors; when the number of existing high-fidelity samples is less than... season ;

[0167] Maximum divergence of the proxy model Predicting the PDE residuals of physical fields and decision space distance Normalization was performed separately to obtain the maximum divergence degree of the normalized surrogate model. Normalized prediction physical field PDE residuals and normalized decision space distance ;

[0168] Among them, for any continuous index The normalization process is as follows:

[0169] ;

[0170] in, express , or , and These represent the minimum and maximum values ​​of the corresponding indicators in the current set of low-confidence candidate samples, respectively.

[0171] For example, after generating massive amounts of augmented sample data with pseudo-labels using a Tri-training semi-supervised learning mechanism in the inner loop, to completely prevent the accumulation of minor physical errors that may be caused by pseudo-labels with iterations, which could lead to catastrophic deviations in the spatiotemporal inference agent's predictions, this application introduces an agent set query strategy based on active learning as an outer loop control mechanism. This outer loop control mechanism utilizes a particle swarm optimization (PSO) algorithm to perform targeted searches within the vast well control decision space, accurately triggering a high-fidelity full-physics numerical simulator to achieve agent calibration under limited computing power. In the reservoir production candidate schemes, the maximum divergence (i.e., the blind zone with the highest uncertainty and most ambiguous physical field) of the prediction results of the three spatiotemporal inference agents is calculated as follows: Only The most uncertain scheme is input into the numerical simulator for precise calculation to obtain the absolutely real development response, and then fed back into the sample dataset. With a very small number of full reservoir numerical simulator calls, the evaluation bias of the spatiotemporal inference agent in high-dimensional space can be significantly reduced, greatly improving the evaluation accuracy and robustness of the spatiotemporal inference agent set.

[0172] This high-cost numerical simulation calibration based on active learning proxy set queries for the outer loop does not invoke a full physical numerical simulator for all reservoir production candidate schemes. Instead, it uses the prediction discrepancies generated by the spatiotemporal inference proxy set during the optimization iteration process as the criterion for whether real numerical simulation calibration is needed. Specifically, in the... In the round of optimization iterations, a set of candidate injection-production parameter vectors is generated by the multi-objective evolutionary algorithm, resulting in the following set of candidate reservoir production schemes:

[0173] ;

[0174] in, This represents the set of candidate solutions for reservoir production. Indicates the first One reservoir production candidate scheme.

[0175] For any reservoir production candidate scheme in the reservoir production candidate scheme set, the corresponding maximum divergence degree is obtained. When the maximum divergence degree is small, it indicates that the prediction results of the three spatiotemporal inference agents for this injection-production scheme are highly consistent, and the scheme can be directly evaluated by the spatiotemporal inference agents. When the maximum divergence degree is large, it indicates that the scheme is in a high uncertainty region where the spatiotemporal inference agents have not yet been fully learned. If the prediction is still relied on only by the spatiotemporal inference agents, the multi-objective optimization algorithm may be misled by incorrect objective function values. Therefore, the embodiments of this application will... As an important indicator for triggering high-fidelity calibration.

[0176] When high-fidelity calibration is triggered, the outer loop high-fidelity calibration sample can be defined as:

[0177] ;

[0178] in, The high-fidelity calibration sample is represented by argmax, which indicates the search for the maximum value. This means prioritizing reservoir production candidate schemes with large discrepancies, large physical residuals, proximity to the Pareto front, and distance from existing high-fidelity samples for full physical numerical simulation. In this way, the high-fidelity numerical simulator is no longer used as a routine evaluation tool for each reservoir production candidate scheme, but is instead triggered at specific points as a truth calibrator for spatiotemporal inference.

[0179] For the selected Call the reservoir numerical simulator to obtain the actual development response:

[0180] ;

[0181] in, This indicates a genuine development response. This represents the simulation process of the reservoir numerical simulator. The actual development response can include actual oil production, water production, gas production curves, pressure field, saturation field, etc. and Wait for the results. However, it is worth noting that the actual development response may contain a lot of data; in this embodiment, only the data that is the same as the output data of the spatiotemporal inference agent needs to be taken.

[0182] Therefore, based on reservoir production candidate schemes and their corresponding real development responses, expanded and calibrated sample training data can be constructed and added to the sample dataset. Then, the spatiotemporal inference agent can be further trained using this sample dataset with the expanded and calibrated sample training data, thus completing the outer loop closure of one optimization candidate scheme generation, agent divergence identification, a small amount of high-fidelity calibration, and agent set update. Through this mechanism, the reservoir numerical simulator is only invoked at key sample locations, avoiding the computational burden of high-cost simulation for each reservoir production candidate scheme in traditional evolutionary optimization, while ensuring that the spatiotemporal inference agent does not experience error drift due to long-term use of pseudo-labels.

[0183] In one possible implementation, based on the predicted reservoir production index data, the objective function values ​​of economic parameters and cumulative oil production are obtained. Then, based on these objective function values, the reservoir production candidate schemes are subjected to non-dominated ranking, crowding distance screening, and genetic operations to obtain updated reservoir production candidate schemes, including:

[0184] Based on the predicted reservoir production index data, the objective function values ​​for economic parameters and cumulative oil production are obtained as follows:

[0185] ;

[0186] ;

[0187] in, This represents the objective function value of the economic parameters. This represents the objective function value of cumulative oil production. Indicates total production time steps; This represents the length of the nth time step. Indicates the discount rate. Indicates oil price, Indicates the price of water treatment. Indicates the total number of producing wells. Indicates the total number of injection wells. This represents the oil production of the j-th producing well in the nth time period. This represents the water production of the j-th production well in the nth time period. Indicates the price of water injection. Let i represent the water injection volume of the j-th injection well in the n-th time period, and i represent the injection well.

[0188] Based on the objective function values ​​of the economic parameters and the objective function value of the cumulative oil production, the reservoir production candidate schemes are subjected to non-dominated sorting, crowding distance screening, and genetic operations to obtain updated reservoir production candidate schemes.

[0189] For example, the spatiotemporal inference proxy set, trained with both inner loop expansion and outer loop calibration, is used as the evaluation engine and seamlessly coupled into the Non-Dominated Ranking Genetic Algorithm (NSGA-II). With net present value (NPV) and cumulative oil production (COP) over the entire lifecycle as dual objectives, the final output response f of the proxy set is... ens (x) Using a weighted summation of the three models, the Pareto optimal injection and extraction parameter set that balances economic benefits and swept volume is output at a speed of milliseconds. The multi-objective optimization model for optimizing variables, design space, objective function, and production scheme is as follows:

[0190] ;

[0191] ;

[0192] Will As the first objective function, Therefore, based on the settings of optimization variables, design space, and objective function, a multi-objective optimization model for the production scheme is established as follows:

[0193] ;

[0194] in, Denotes the first objective function. This represents the second objective function. This represents a multi-objective optimization model for production plans. This indicates candidate solutions for reservoir production.

[0195] In this invention, the multi-objective evolutionary algorithm does not directly call the reservoir numerical simulator to calculate the objective function value of each candidate injection and production scheme. Instead, it uses a sample dataset with expanded and calibrated sample training data and a spatiotemporal inference agent set as a fast evaluation engine.

[0196] For example, embodiments of this application can employ NSGA-II (NSGA-II is a fast, elitist, multi-objective genetic algorithm) for iteration. NSGA-II first generates the first... t The generation population is:

[0197] ={ , ,…, ,…, };

[0198] in, Let represent the k-th reservoir production candidate scheme in the t-th iteration. For traditional methods, each... The objective function value can only be obtained by inputting into a reservoir numerical simulator; however, in the embodiments of this application, each... First, the input spatiotemporal inference agent set is used for millisecond-level prediction, and then the response results output by the spatiotemporal inference agent set are further transformed into two optimization objectives by the objective function calculation module:

[0199] ;

[0200] ;

[0201] in, Indicates the approval of reservoir production candidate schemes The net present value obtained, Indicates the approval of reservoir production candidate schemes The cumulative oil production obtained, This represents the value of the first objective function. This represents the value of the second objective function.

[0202] Assuming there are candidate reservoir production schemes Candidate solutions for reservoir production And satisfy:

[0203] ;

[0204] ;

[0205] At the same time, at least one of the following conditions must be met:

[0206]

[0207] or:

[0208]

[0209] Then determine the candidate reservoir production schemes. Candidate schemes for controlling reservoir production .in, and These represent candidate reservoir production schemes. and The corresponding predicted net present value objective function value; and These represent candidate reservoir production schemes. and The corresponding objective function value for the predicted cumulative oil production; and These represent the preset net present value advantage threshold and the preset cumulative oil production advantage threshold, respectively.

[0210] Based on this dominance relationship, NSGA-II targets the current population. A non-dominated sorting process is performed to obtain multiple non-dominated levels. Candidate schemes are retained sequentially from highest to lowest non-dominated level. When the number of candidate schemes in a certain non-dominated level exceeds the remaining capacity of the next generation population, the crowding distance of each candidate scheme within that non-dominated level is calculated. Candidate schemes with larger crowding distances are then selected and retained in descending order until the next generation population reaches the preset population size. Subsequently, selection, crossover, and mutation operations are performed sequentially on the retained candidate schemes to generate a set of next generation candidate injection / collection schemes.

[0211]

[0212] in, Indicates the first Collection of candidate solutions for reservoir production Indicates the first Collection of candidate solutions for reservoir production; Indicates the distance from the non-dominated hierarchy and the crowding distance from the first In the process of selecting parent candidate schemes in the generation population, priority is given to candidate schemes with higher non-dominated levels and larger crowding distances. This indicates an operation of cross-recombining the well control decision variables in the parent candidate scheme to generate new candidate injection-production parameter combinations; This refers to the operation of randomly perturbing some well-control decision variables in the candidate schemes after crossover, in order to increase population diversity and avoid the optimization process from getting trapped in local optima.

[0213] Newly generated The trained spatiotemporal inference agent set is then quickly evaluated to obtain the corresponding objective function values ​​for predicted net present value and predicted cumulative oil production. The non-dominated sorting, crowding distance screening, and genetic operations described above are then repeated. Thus, the spatiotemporal inference agent set participates in optimization as a fast objective function evaluator in each NSGA-II iteration, while the reservoir numerical simulator performs high-fidelity calibration only on a small number of candidate schemes with high uncertainty or high calibration value. This reduces the frequency of reservoir numerical simulator calls, improves optimization efficiency, and reduces the black-box bias of the purely data-driven agent model through physical constraints and active calibration.

[0214] In an optional implementation, to prevent the spatiotemporal inference agent from deviating from the true numerical simulation response during long-term evolutionary search, this embodiment sets a dynamic closed-loop calibration mechanism in the NSGA-II iteration. When any of the following calibration conditions are met, a high-value scheme is selected from the current reservoir production candidate schemes for fidelity numerical simulation calibration. The calibration conditions are:

[0215] Condition 1: ;

[0216] Condition 2: ;

[0217] in, This represents the threshold of agent uncertainty. Indicates the physical residual threshold. Indicates candidate reservoir production schemes The corresponding PDE residuals for predicting the physical field.

[0218] If the variation of the current Pareto front is less than the set threshold for several consecutive generations, but the divergence of the surrogate set is still high, a high-value scheme can be selected from the current reservoir production candidate schemes for high-fidelity calibration.

[0219] Therefore, the spatiotemporal inference agent in this embodiment is not only used for offline prediction before optimization, but is embedded in every population evaluation process of NSGA-II as a fast calculation engine for the objective function value and constraint state of reservoir production candidate schemes. Simultaneously, the spatiotemporal inference agent maintains periodic calibration with the high-fidelity reservoir numerical simulator through an outer-loop active learning mechanism. This coupling method enables NSGA-II to complete tens of thousands of reservoir production candidate scheme evaluations without frequently calling the full-physics numerical simulator, and ultimately outputs the NPV-COP bi-objective Pareto optimal injection-production parameter set. Compared with multi-objective optimization driven by direct numerical simulation, this embodiment transforms the high-cost numerical simulator from a fitness calculator that must be called for each individual into a calibrator that is only called for key high-uncertainty samples, thereby significantly reducing computational costs while ensuring traceable physical consistency and high-fidelity reliability of the optimization results.

[0220] Based on the above technical solution, let's take a typical porous, thick, unfractured carbonate rock block in the Middle East as an example. This block has three dimensions of 2000m × 1800m × 150m (thickness), and the total number of grids reaches 3.5 million; For example... Figure 2 As shown, it includes 20 oil production wells (with constant bottom pressure) and 10 water injection wells (with controlled water injection volume), with an optimization cycle of 20 steps and 600 decision variables.

[0221] Implementation Process and Results: Initially, the system constructed its initial sample using only 300 sets of real simulation samples obtained through Latin hypercube sampling (LHS). In the subsequent 50,000 NSGA-II population propagation runs, the inner-loop Tri-training mechanism generated 15,000 high-confidence samples with pseudo-labels at zero computational cost and corrected them using the PINN equation; while the outer-loop high-cost numerical simulation calibration mechanism only triggered 50 full-reservoir numerical simulations when model divergence was extremely high. Finally, as... Figure 3 As shown, compared with conventional optimization based on a single proxy model, the Pareto optimal NPV scheme selected in this invention improves the water drive sweep efficiency of the block by 12%, increases the cumulative oil production by 4.5%, and keeps the material balance error within 1%, which proves its high efficiency in the optimization of the development of large and complex oil reservoirs under limited computing power.

[0222] The production scheme with the highest NPV is selected as Scheme 1, and the production scheme with the highest cumulative oil production is selected as Scheme 2. The optimized results are then demonstrated. Figure 4 (a) shows the production wellbore pressure of Scheme 1. Figure 4 (b) shows the production wellbore pressure of Scheme 2; Figure 5 (a) shows the injection volume of the injection well in Scheme 1. Figure 5 (b) shows the injection volume of the injection well in Scheme 2; Figure 4(a) and Figure 4 In (b), the parameter axis 0-15 represents pressure, with units of MPa; Figure 5 (a) and Figure 5 In (b), the parameter axis from 0 to 1500 represents the injection volume, in meters (m). 3 / sky; Figure 6 The NPV and cumulative oil production curves for Scheme 1 and Scheme 2 are shown.

[0223] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0224] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0225] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0226] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0227] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0228] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing production in porous, thick carbonate reservoirs, characterized in that, include: Sample geological well control parameters and corresponding ground value label data of the sample geological well control parameters are collected during the production process of porous, thick carbonate reservoirs to obtain sample training data; the ground value label data represents reservoir production index data. Construct a set of spatiotemporal inference agents and a matter conservation loss function coupled with a gravity term; the set of spatiotemporal inference agents includes at least one spatiotemporal inference agent; Based on the sample training data and the matter conservation loss function, the spatiotemporal inference agent set is trained to obtain the trained spatiotemporal inference agent set. Generate reservoir production candidate schemes, and use the trained spatiotemporal inference agent set to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data. Based on the predicted reservoir formation index data, the sample training data is expanded by using inner-loop zero-cost sample expansion based on Tri-training collaborative training and outer-loop high-cost numerical simulation calibration based on active learning agent set query, and the expanded sample training data is obtained. Based on the predicted reservoir production index data, the objective function values ​​of economic parameters and cumulative oil production are obtained. Then, based on the objective function values ​​of economic parameters and cumulative oil production, non-dominated sorting, crowding distance screening, and genetic operations are performed on the reservoir production candidate schemes to obtain updated reservoir production candidate schemes. The process involves iterating repeatedly using expanded and calibrated sample training data and updated reservoir production candidate schemes until the iteration termination condition is met. Then, based on the reservoir production candidate schemes after non-dominated sorting in the last iteration, the target reservoir production scheme for porous, thick carbonate reservoirs is determined.

2. The method for optimizing production in porous, thick carbonate reservoirs according to claim 1, characterized in that, The sample geological well control parameters include a three-dimensional static geological attribute field, effective reservoir thickness, the ratio of vertical to horizontal permeability characterizing vertical heterogeneity, and a vector of well control decision variables.

3. The method for optimizing production in porous, thick carbonate reservoirs according to claim 1, characterized in that, Construct a spatiotemporal inference agent set, including: A first, second, and third spatiotemporal inference agent, all of which are CNN-LSTM models, are constructed to obtain a spatiotemporal inference agent set. The output of the spatiotemporal inference agent set is obtained by weighted summation of the outputs of the first, second, and third spatiotemporal inference agents.

4. The method for optimizing production in porous, thick carbonate reservoirs according to claim 1, characterized in that, The mass conservation loss function for the coupled gravity term is constructed as follows: ； ； ； ； ； ； in, The mass conservation loss function representing the coupled gravity term. This represents the mean square error term. Indicates the weight of the data item. Indicates the physical constraint penalty weight. This indicates the number of spatiotemporal coordinate points within the computational domain. This represents the partial differential parameter of the oil phase at the z-th spatiotemporal collocation point. This represents the partial differential parameter of the oil phase at the z-th spatiotemporal collocation point. The function representing the calculation of the partial differential parameters of the aqueous phase. The function representing the calculation of the partial differential parameters of the oil phase. Represents the differential symbol. Indicates porosity. This represents the density of the aqueous phase. Indicates time, Represents the gradient. Indicates the water phase flow rate. Indicates oil phase flow rate, Represents the velocity vector of water phase seepage. Indicates the density of the oil phase. This represents the oil phase flow velocity vector. The function representing the calculation of the velocity vector of water phase seepage. The function representing the calculation of the oil phase seepage velocity vector. Indicates spatial location The absolute permeability tensor at that location, This indicates the spatial location of spatiotemporal points or grid centers within the three-dimensional computational domain of porous, thick carbonate reservoirs. Indicates the relative permeability of water. Indicates the viscosity of the aqueous phase. Indicates spatial location and time The corresponding predicted pressure field value, Represents spatial coordinate vectors The gradient operator, where g represents gravitational acceleration. This represents the spatial gradient along the direction of gravity. Indicates the relative permeability of oil. Indicates the viscosity of the oil phase. This represents the predicted value of the oil saturation field. This represents the predicted value of the water saturation field.

5. The method for optimizing production in porous, thick carbonate reservoirs according to claim 1, characterized in that, Based on the sample training data and the matter conservation loss function, the spatiotemporal inference agent set is trained to obtain the trained spatiotemporal inference agent set, including: Based on the sample training data and with the goal of minimizing the matter conservation loss function, the spatiotemporal inference agent set is trained using the gradient descent algorithm to obtain the trained spatiotemporal inference agent set.

6. The method for optimizing production in porous, thick carbonate reservoirs according to claim 2, characterized in that, Generate reservoir production candidate schemes, and use the trained spatiotemporal inference proxy set to analyze and predict the reservoir production candidate schemes to obtain predicted reservoir production index data, including: Generate reservoir production candidate schemes; the reservoir production candidate schemes include a vector of candidate well control decision variables; The three-dimensional static geological property field, effective reservoir thickness, and vertical to horizontal permeability ratio characterizing vertical heterogeneity were collected during the production process of porous, thick carbonate reservoirs, and these were combined with candidate reservoir production schemes to form real-time geological well control parameters. The real-time geological well control parameters are used as input to the trained spatiotemporal inference agent set, and the predicted reservoir production index data output by the trained spatiotemporal inference agent set are obtained. The predicted reservoir production index data includes the predicted reservoir production index output by the first spatiotemporal inference agent, the second spatiotemporal inference agent, the third spatiotemporal inference agent, and the spatiotemporal inference agent set.

7. The method for optimizing production in porous, thick carbonate reservoirs according to claim 6, characterized in that, Based on the predicted reservoir formation index data, the sample training data is expanded using a zero-cost inner-loop sample expansion based on Tri-training collaborative training and a high-cost outer-loop numerical simulation calibration based on active learning agent set query, resulting in expanded sample training data, including: Based on the predicted reservoir generation index data, the maximum divergence of the reservoir production candidate schemes is evaluated, and the maximum divergence of the reservoir production candidate schemes is obtained as follows: ； in, This represents the maximum degree of divergence corresponding to the expanded sample data. This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... Predicted reservoir production indicators obtained by a spatiotemporal inference agent; This indicates that the spatiotemporal inference proxy input data x, constructed based on reservoir production candidate schemes, is used to... The predicted reservoir production indicators obtained by the spatiotemporal inference agent; max represents the maximum value function. =1,2,3; =1,2,3; A high-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is less than a pre-set divergence threshold. A low-confidence candidate sample set is constructed by selecting reservoir production candidate schemes whose maximum divergence is greater than or equal to a pre-set divergence threshold. Based on the high-confidence candidate sample set, perform zero-cost inner-loop sample expansion using Tri-training collaborative training, and perform high-cost outer-loop numerical simulation calibration based on active learning agent set query using the low-confidence candidate sample set to obtain expanded sample data; The original sample training data is expanded using the expanded sample data to obtain the expanded sample training data.

8. The method for optimizing production in porous, thick carbonate reservoirs according to claim 7, characterized in that, Based on the high-confidence candidate sample set, perform inner-loop zero-cost sample augmentation using Tri-training co-training, including: For any high-confidence reservoir production candidate scheme in the high-confidence candidate sample set, obtain the mean value of the predicted reservoir production index corresponding to the predicted reservoir production index output by the two spatiotemporal inference agents that participated in the calculation of the maximum divergence degree; Based on the high-confidence reservoir production candidate schemes, sample geological well control parameters are constructed, and the average predicted reservoir production indexes are used to construct corresponding pseudo-label data to obtain the first expanded sample data.

9. The method for optimizing production in porous, thick carbonate reservoirs according to claim 8, characterized in that, High-cost numerical simulation calibration of the outer loop based on active learning agent set queries, using a set of low-confidence candidate samples, includes: For any low-confidence reservoir production candidate scheme in the set of low-confidence candidate samples, obtain the prediction physics field (PDE) residual corresponding to the low-confidence reservoir production candidate scheme. Pareto Front Proximity Indicator Function and the decision space distance with existing high-fidelity samples. ; This represents a low-confidence reservoir production candidate scheme in the low-confidence candidate sample set; the existing high-fidelity sample refers to the existing second augmented sample data; The maximum degree of divergence corresponding to the production candidate schemes for the low-confidence reservoir. Predicting the PDE residuals of physical fields and decision space distance Normalization was performed separately to obtain the maximum divergence degree of the normalized surrogate model. Normalized prediction physical field PDE residuals and normalized decision space distance ; Based on the maximum divergence of the normalized agency model Normalized prediction physical field PDE residuals Normalized decision space distance and the Pareto front proximity indicator function Calculate the production candidate schemes for the low-confidence reservoir. The high-fidelity calibration value function value is: ； in, Indicates low-confidence reservoir production candidate schemes High-fidelity calibration value function value. , , , Let these represent the first weight, second weight, third weight, and fourth weight, respectively, and satisfy the following conditions: , ; According to the high-fidelity calibration value function value The low-confidence reservoir production candidate schemes in the low-confidence candidate sample set are sorted from largest to smallest, and the schemes ranked first are selected. Low-confidence reservoir production candidate schemes are used as calibration samples; or schemes that meet the following criteria are selected. Low-confidence reservoir production candidate schemes were used as calibration samples; among them, This indicates the maximum number of samples allowed to be called from the reservoir numerical simulator in each iteration. This indicates a preset high-fidelity calibration value threshold; The reservoir numerical simulator is called to perform simulation calibration on the sample to be calibrated to obtain the corresponding real reservoir production index data. The geological well control parameters of the sample to be calibrated are constructed, and the real reservoir production index data corresponding to the sample to be calibrated are used as the corresponding high-fidelity label data to obtain the second expanded sample data.

10. The method for optimizing production in porous, thick carbonate reservoirs according to claim 1, characterized in that, Based on the predicted reservoir production index data, the objective function values ​​of economic parameters and cumulative oil production are obtained. Then, based on these values, non-dominated ranking, crowding distance screening, and genetic operations are performed on the reservoir production candidate schemes to obtain updated reservoir production candidate schemes, including: Based on the predicted reservoir production index data, the objective function values ​​for economic parameters and cumulative oil production are obtained as follows: ； ； in, This represents the objective function value of the economic parameters. This represents the objective function value of cumulative oil production. Indicates total production time steps; This represents the length of the nth time step; Indicates the discount rate. Indicates oil price, Indicates the price of water treatment. Indicates the total number of producing wells. Indicates the total number of injection wells. This represents the oil production of the j-th producing well in the nth time period. This represents the water production of the j-th production well in the nth time period. Indicates the price of water injection. Let i represent the water injection volume of the j-th injection well in the n-th time period, and i represent the injection well. Based on the objective function values ​​of the economic parameters and the objective function value of the cumulative oil production, the reservoir production candidate schemes are subjected to non-dominated sorting, crowding distance screening, and genetic operations to obtain updated reservoir production candidate schemes.

Images