A robust optimization method for water-drive reservoirs based on data-driven physical models

Abstract

This invention discloses a robust optimization method for water-drive reservoirs based on a data-driven physical model, relating to the field of oil and gas field development technology. The invention first divides the flow control grid according to the inter-well connectivity topology of the water-drive reservoir, constructs a block-shaped capacitance-resistance physical model, determines the multi-level tensor iterative equations for multi-stage fluid series transmission, and constructs a priori dataset based on reservoir physical priors and statistical characteristics of water injection conditions. Then, a generative adversarial network (GAN) is constructed, and the priori dataset is input into it to obtain multiple sets of CRM-Block attribute parameter fields, generating a set of reservoir uncertainty scenarios. Using the decision variables in the reservoir injection-production scheme as the optimization objective, a feasible region constraint space is set, and a robust optimization model is established. The robust optimization model is solved in the feasible region constraint space using a particle swarm optimization algorithm to obtain the optimal robust injection-production scheme. This invention achieves low-cost acquisition of the optimal robust injection-production scheme for water-drive reservoirs, which is beneficial for guiding the injection-production development of water-drive reservoirs.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development technology, and specifically to a robust optimization method for water-drive reservoirs based on a data-driven physical model. Background Technology

[0002] The efficient development of water-driven oilfields relies heavily on the continuous optimization of injection and production schemes. This optimization inevitably depends on mathematical models capable of accurately predicting reservoir dynamics. Traditional reservoir numerical simulation methods possess rigorous physical mechanisms and precise calculations, but their geological modeling processes are exceptionally complex, and history fitting and prediction calculations are extremely time-consuming, making it difficult to meet the urgent needs of modern intelligent oilfields for high-frequency, real-time optimization decision-making. As shown, employing surrogate models to assist or even replace traditional reservoir numerical simulations has become a key technological development direction in this field.

[0003] The Capacitor-Resistant Model (CRM model) is a typical surrogate model that combines clear physical meaning with extremely high computational efficiency. However, as a typical lumped-parameter model, the classic CRM model usually simplifies the fluid flow region between injection and production wells into a single ideal container, using only a single connectivity coefficient and inter-well constant to characterize the dynamic relationship between wells. This makes it extremely limited when applied to water-injected reservoirs with complex porosity and seepage evolution and heterogeneous flow channels. This single-source, single-sink zero-dimensional assumption completely ignores the spatial distribution characteristics of fluid along streamlines in porous media, and cannot accurately characterize the multi-stage hysteresis effects and dynamic attenuation laws during fluid transport. With the deepening of reservoir water injection development, the actual flow of fluid between wells often exhibits strong spatiotemporal heterogeneity. Traditional single-node CRM models struggle to capture fine dynamic characteristics and cannot provide a high-resolution physical feature space for subsequent intelligent algorithms.

[0004] To improve physical fidelity, the introduction of a block capacitance-resistance model (CRM-Block module) that discretizes the inter-well flow channel into multiple series-connected response units has become an inevitable trend. However, before using the CRM model for production forecasting, it is necessary to obtain the core parameters of the model through reservoir history fitting and inversion. With the multi-level spatial expansion of the model, the dimensionality of the parameters to be inverted increases dramatically. This high-dimensional parameter inversion has strong mathematical ambiguity, meaning that there are many drastically different parameter combinations that can highly match the actual reservoir historical observation data. This ambiguity brings fundamental uncertainty to the future prediction of the reservoir, easily leading to the problem of a good fit but inaccurate prediction.

[0005] As our understanding of decision-making risks deepens, random sampling methods are being gradually introduced to explore all possible solutions for parameters, thus fully characterizing uncertainty. However, these methods suffer from a fatal flaw: computational efficiency. When faced with complex, multi-layered CRM models with exponentially increasing parameter dimensions, random sampling methods require massive iterative sampling, leading to exponentially increasing computational costs and hindering timely support for on-site decision-making. Furthermore, existing reservoir decision-making methods have consistently failed to effectively reconcile the significant contradiction between decision reliability and computational timeliness.

[0006] Therefore, there is an urgent need to propose a robust optimization method for water-drive reservoirs based on a data-driven physical model, which takes into account both the characterization of complex geological features and the efficient and intelligent generation of uncertain parameter spaces, so as to achieve the low-cost and robust acquisition of the optimal injection and production scheme for water-drive reservoirs. Summary of the Invention

[0007] This invention aims to solve the above-mentioned problems and proposes a robust optimization method for water-drive reservoirs based on a data-driven physical model. It takes into account both the detailed characterization of complex geological features and the efficient and intelligent generation of the uncertainty parameter space. It solves the problem of multiple solutions in CRM-Block parameter inversion and the shortcomings of traditional reservoir injection and production scheme optimization methods, such as the inability to quantify uncertainty and the poor robustness of optimization results. It realizes the low-cost acquisition of the optimal robust injection and production scheme for water-drive reservoirs and provides a basis for guiding the injection and production development of water-drive reservoirs.

[0008] The present invention adopts the following technical solution: A robust optimization method for water-drive reservoirs based on a data-driven physical model includes the following steps: Step 1: Obtain the injection and production scheme of the water-driven reservoir in the target block, divide the flow control grid according to the inter-well connectivity topology, construct the block capacitance and resistance physical model, determine the multi-level tensor iterative equation for multi-level series transmission of fluid, and construct the prior dataset based on reservoir physical priors and statistical characteristics of water injection conditions. Step 2: Construct a generative adversarial network and input the prior samples in the prior dataset into the generative adversarial network to map and generate multiple sets of CRM-Block attribute parameter fields; Step 3: Generate a set of reservoir uncertainty scenarios based on the CRM-Block attribute parameter field, and set a feasible domain constraint space that satisfies the actual production and operation of the reservoir, using the decision variables in the reservoir injection and production scheme as the optimization objective. Furthermore, a robust optimization model was established to seek robust injection and production schemes for reservoirs under the worst geological conditions. Step 4: Utilize the particle swarm optimization algorithm within the feasible region constraint space that satisfies the actual production and operation of the reservoir. Solve the robust optimization model to obtain the optimal robust injection and extraction scheme.

[0009] Preferably, step 1 includes the following sub-steps: Step 1.1: Based on the injection and production plan for the water-drive reservoir, obtain the total number of water injection wells for the water-drive reservoir. Total number of production wells The topological relationship between wells is defined by the streamline trajectory from the injection well to the production well. A series of flow control grids are used to characterize the multi-level delay effects of pressure signals and fluid transport in water-drive reservoirs, and a connectivity coefficient matrix is ​​constructed. and Group time constant matrix The physical model of a block capacitor and resistor, wherein, , , For the set of real numbers, This refers to the serial number of the flow control grid. , This represents the total number of serial flow control grids; Step 1.2: Based on the assumption of a pure water-driven steady-state oil reservoir and ignoring the primary oil recovery effect, determine the multi-stage tensor iterative equation for multi-stage series transmission of fluid according to the discretized tensor forward transfer equation, including the flow response equations of the first-stage grid and other-stage grids in the series flow control grid, which are used to predict the total production matrix of the production well. Step 1.3: Based on the injection and production scheme of the water-drive reservoir in the target block, extract the statistical characteristics of the water injection conditions as a priori template. In the time dimension, dynamically set the allocation mutation nodes based on a stochastic process. In the numerical dimension, introduce global and single-well independent scaling factors to nonlinearly perturb the priori template to generate a stepped water injection sequence matrix. ; Step 1.4: Construct physical parameter field labels, including a sparsed connectivity coefficient matrix. The set of time constant matrices that follow a log-normal distribution The generated stepped water injection sequence matrix Substituting the physical parameter field labels into the multi-level tensor iterative equation, and superimposing a random Gaussian noise matrix positively correlated with its absolute magnitude onto the total liquid production matrix, yields the synthetic liquid production sequence with measurement errors. The synthetic product sequence Stepped water injection sequence matrix Prior samples are generated by combining the physical parameter field labels with the prior samples to construct a prior dataset.

[0010] Preferably, in the multi-stage tensor iterative equation for multi-stage fluid transmission, the flow response equation is determined for the first-stage grid in the series flow control grid as follows: ; In the formula, for The predicted output flow tensor of the first-level grid at time step; for The predicted output flow tensor of the first-level grid at time step; The product of Hadama; It is a natural exponential function; This represents the discretization time step. The time constant matrix of the first-level grid; for The vector of actual water injection rates of each injection well in a water-driven reservoir at any given time; It is a matrix of all 1s; For the other levels of the cascaded flow control grid besides the first-level grid, the flow response equation is determined as follows: ; In the formula, for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid.

[0011] Preferably, the generative adversarial network includes a generator, a differentiable physical layer, and a discriminator; The generator includes an input fusion layer, a temporal feature encoder, a self-attention module, and a multi-branch parameter decoder. The input fusion layer uses the actual water injection sequence of the water-drive reservoir. With actual liquid production sequence The injection-sampling time series is formed by concatenating the data along the feature channel dimension, and a random noise vector sampled from a standard normal distribution is introduced. Construct a multivariate input conditional feature matrix The temporal feature encoder contains five one-dimensional convolutional neural networks to extract multi-scale local dynamic features and obtain deep convolutional feature maps. The self-attention module, by connecting a multi-head self-attention layer after the deep convolutional feature maps, is used to capture long sequence dependencies. The self-attention module calculates scaled dot product attention weights and outputs context fusion features. The multi-branch parameter decoder is used to process context fusion features. The output generator outputs parameters and internally has a first branch and a second branch. The first branch obtains the connectivity coefficient matrix through the sigmoid activation function. The second branch obtains the time constant matrix through the Softplus activation function. Form a set of multi-level time constant matrices According to the connectivity coefficient matrix and multi-level time constant matrix set Output Generator Output Parameters ; The differentiable physical layer is used to substitute the generator output parameters into the multi-level tensor iterative equation for forward modeling, and backpropagate the predicted output flow tensor obtained from the forward modeling to the generator to achieve fully end-to-end training. The discriminator is built on a least-squares adversarial architecture. First, the injected time series is input into three sequentially connected one-dimensional convolutional layers for downsampling, then input into an adaptive average pooling layer. After three rounds of one-dimensional convolutional downsampling and pooling compression, the data is flattened to obtain the temporal conditional feature vector. This is used to extract deterministic high-dimensional hidden features from the dynamic injection-production history of water-drive reservoirs within the target block; then the generator output parameters are... Flatten the input and map it to a fully connected layer as a static parameter feature vector. This is used to extract high-dimensional hidden features characterizing the topological structure of underground fluid connectivity; then a discriminator is used to compare the temporal conditional feature vectors. With static parameter eigenvectors The feature matching degree is used to determine whether the feature vector is true or false, and the temporal conditional feature vector is used for this purpose. With static parameter eigenvectors By splicing along the channel dimension, scalar scores are generated by outputting single-neuron linear layers without activation functions through a multilayer perceptron, and then mapped to generate multiple sets of CRM-Block attribute parameter fields.

[0012] Preferably, the total loss function of the generative adversarial network Set to: ; in, ; ; ; In the formula, The hyperparameter weighting coefficients are used to counteract the loss term; To counteract the loss; These are the hyperparameter weighting coefficients for the physical response loss term; The physical response loss term is used to constrain the generated physical parameter field to accurately fit the actual historical liquid production rate after being substituted into the differentiable physical layer for forward modeling. These are the hyperparameter weight coefficients for the parameter boundary constraint loss term; For parameter boundary constraint loss terms; For the weighting coefficients of the hyperparameters of the multi-solution diversity regularization loss term; This is a loss term for multi-solution diversity regularization, used to penalize random noise vectors. Mode collapse behavior where changes in physical parameters, field parameters, and structural changes do not alter the generated physical parameters; For mathematical expectation operations; For random noise vectors It follows a standard normal distribution; The linear confidence score is the score without the Sigmoid mapping. For generator functions, , For injection and extraction conditions, It is a random noise vector. Let be the connectivity coefficient matrix. It is a set of multi-level time constant matrices. The time constant matrix; The linear confidence score is the score without the Sigmoid mapping. This represents the total number of time steps. This refers to the time step number; For production well serial numbers; Total number of producing wells; for Time of the first Predicted fluid production of a production well; for Time of the first The actual fluid production of a production well; It is a function for maximizing the value; It is an L1 norm; The preset disturbance response threshold is used to limit the minimum difference in the physical parameter fields generated by different random noise vectors. , All of these are generated physical parameter fields. , , , All of them are independent random noise vectors.

[0013] Preferably, step 3 includes the following sub-steps: Step 3.1: Set the total production forecast period for future reservoir optimization. and according to time intervals Discretely divide it The injection volume of each water injection well in the reservoir during each control period of equal duration is used as a decision variable to construct an injection volume decision; The generated multiple sets of CRM-Block attribute parameter fields are used as discrete reservoir uncertainty scenarios to construct a reservoir uncertainty scenario set. , ,in, This represents the sequence number of the reservoir uncertainty scenario. , This represents the total number of uncertain scenarios in the oil reservoir. The first in the set of reservoir uncertainty scenarios There are several reservoir uncertainty scenarios, each including a set of connectivity coefficient matrices. as well as Group of multi-level time constant matrices Each reservoir uncertainty scenario corresponds to a different geological dynamic response mode; Step 3.2: For each reservoir uncertainty scenario in the reservoir uncertainty scenario set, obtain the injection volume decision, call the generative adversarial network, and use the differentiable physics layer in the generative adversarial network to perform forward modeling prediction to obtain the total production volume sequence of each production well in the reservoir within each specified control period. , ; Based on the sequence of total production volume of each production well in the reservoir during each specified control period, with the injection volume of water injection wells as the decision variable and the total production volume of production wells as the objective, a comprehensive benefit objective function is established to characterize the energy consumption of water injection and the revenue from production, so as to achieve a quantitative evaluation of the benefits of the reservoir from the injection end to the production end. Step 3.3: Construct a feasible domain constraint space that satisfies the actual production and operation of the reservoir. By transforming robust decision-making theory into a mathematical optimization problem with a minimax game structure to determine robust injection-production schemes for reservoirs, a robust optimization model is constructed to find robust injection-production schemes that maximize the overall reservoir benefit objective function under worst-case reservoir uncertainty scenarios. ; The robust optimization model, based on a preset injection volume decision, traverses all reservoir uncertainty scenarios in the reservoir uncertainty scenario set, identifies the scenario that minimizes the overall benefit objective function, and designates it as the worst-case scenario. Then, it performs optimization within the feasible region constraint space. Robust injection volume decision for maximizing the overall benefit objective function in worst-case scenario for in-situ reservoir searching , ,in, The function is used to calculate the maximum value of the independent variable. It is a minimum value function.

[0014] Preferably, the comprehensive benefit objective function is set as follows: ; in, ; In the formula, The objective function is the comprehensive benefit objective function; For the decision of injection volume, , For the set of real numbers, This represents the total number of injection wells. For the first Within the first control period, the first Injection rate of the injection well; To control the time period sequence number; To control the total number of time periods; For production well serial numbers; Total number of producing wells; This refers to the serial number of the injection well; The economic benefit weight of the unit liquid production volume; Weighting of energy consumption and operating costs per unit volume of water injected; For the first Under the first reservoir uncertainty scenario, the first Predicted production rate of a production well; For the first Within the first control period, the first Injection rate of the injection well.

[0015] Preferably, a feasible domain constraint space that satisfies the actual production and operation of the reservoir is constructed based on constraints such as single-well ultimate injection capacity, water injection capacity, production well fluid throughput constraints, and on-site operational stability constraints. The single-well limit injection capacity constraint is limited by the rated power of the reservoir production equipment and the formation fracture pressure; the water injection capacity constraint is set to the total injection volume of all water injection wells in the reservoir within a single control period not exceeding the maximum water supply capacity of the gathering and transportation station; the production well fluid throughput constraint is set to the predicted production fluid volume not exceeding the maximum daily fluid discharge of the production well lifting equipment; the field operation stability constraint is set to the adjustment range of the injection volume of each water injection well in the reservoir during adjacent control periods must be limited within the proportional boundary of safe fluctuation, and the injection volume of each water injection well in the reservoir during the first control period must be smoothly connected with the actual water injection state at the end of the reservoir's production history.

[0016] Preferably, step 4 includes the following sub-steps: Step 4.1, Set the total number of particle swarm examples. and maximum number of generations The decision on the dispensing quantity to be optimized is expanded and mapped to the particle positions in a high-dimensional search space. Initialize the initial position and initial velocity of each particle in the particle swarm; Step 4.2, utilizing particle swarm optimization in the feasible region constrained space In the process of iterative calculation to solve the robust optimization model, each particle in the particle swarm sequentially traverses each reservoir uncertainty scenario in the reservoir uncertainty scenario set, and calls the generative adversarial network to perform forward modeling to obtain the comprehensive benefit objective function value corresponding to each reservoir uncertainty scenario. The individual fitness of the particle in the current iteration step is determined based on the reservoir uncertainty scenario with the lowest comprehensive benefit objective function value. Step 4.3: Use a robust optimization model to find the particle position that maximizes individual fitness, compare the current individual fitness of each particle with the historical individual fitness record, and update the historical best position of each particle in the particle swarm. And the global optimal position of the entire population. ; Then, determine the individual best position of each particle in the particle swarm. And the global optimal position of the entire population. As a guide, update the velocity and position of each particle in the particle swarm; Step 4.4: Determine if the current iteration step meets the preset iteration termination condition. If the preset iteration termination condition is not met, return to step 4.2 to continue the iterative calculation; otherwise, set the currently determined global optimal position. It then reverse-maps and restores the injection volume decision, outputting the optimal robust injection and production scheme for all reservoir uncertainty scenarios in the reservoir uncertainty scenario set.

[0017] Preferably, in step 4, the formula for calculating the individual fitness of the particle is: ; In the formula, For individual fitness evaluation function; For particle serial numbers; This refers to the iteration step number; For the first During the nth iteration The position of each particle; It is a minimum value function; The objective function is the comprehensive benefit objective function; For the first During the nth iteration Decision on the allocation of individual particles; This represents the sequence number of the reservoir uncertainty scenario. , This represents the total number of uncertain scenarios in the oil reservoir. The first in the set of reservoir uncertainty scenarios Uncertainty scenarios in oil reservoirs; The iteration termination condition is that the current iteration step reaches a preset maximum number of generations. Or the change in the fitness of the individual particle corresponding to the global optimal position is continuous. If the iteration rate is lower than the preset convergence threshold, it is determined that the preset iteration termination condition is met; otherwise, it is determined that the preset iteration termination condition is not met.

[0018] The present invention has the following beneficial effects: This invention proposes a robust optimization method for water-driven reservoirs based on a data-driven physical model. This method solves the limitation of traditional capacitance-resistance models, which cannot accurately characterize complex heterogeneous flow fields and multi-stage hysteresis effects in reservoirs due to the single-node assumption, by constructing a gridded CRM-Block differentiable physical model. Combined with a generative adversarial network based on self-attention mechanism and physical constraints, it achieves effective processing of dynamic injection-production time series data of reservoirs. By transforming the extremely time-consuming traditional reservoir physical parameter inversion into second-level generative sampling, it overcomes the curse of multiple solutions and computational bottleneck of high-dimensional spatial parameter inversion.

[0019] Meanwhile, this invention, relying on the comprehensive benefit objective function of direct fluid mapping, effectively avoids the error accumulation caused by complex water cut prediction. Combined with a built-in robust optimization model to quantify the uncertainty of underground reservoir parameters, it achieves a robust shift in optimization decision-making from "blindly pursuing ideal extreme values" to "effectively guaranteeing minimum returns." This invention not only significantly improves the risk resistance of reservoir injection and production schemes under real geological disturbances but also provides key technical support for high-frequency, real-time dynamic injection and production optimization in smart oilfields. It is conducive to ensuring long-term stable production and maximizing economic benefits in water-driven reservoirs, and has significant engineering application value. Attached Figure Description

[0020] Figure 1 This is a flowchart of a robust optimization method for water-drive reservoirs based on a data-driven physical model, according to the present invention.

[0021] Figure 2 A comparison chart showing the production volume fitting of production well P1.

[0022] Figure 3 A comparison chart showing the production volume fitting of production well P2.

[0023] Figure 4 A comparison chart showing the fluid production of production well P3.

[0024] Figure 5 A comparison chart showing the production volume fitting of production well P4.

[0025] Figure 6 A schematic diagram of the first-order time constant matrix near the water injection end of the water-driven reservoir generated by the generator.

[0026] Figure 7A schematic diagram of the second-order time constant matrix at the deep matrix of a water-drive reservoir generated by the generator.

[0027] Figure 8 A schematic diagram of the third-order time constant matrix near the production end of a water-driven reservoir generated by the generator.

[0028] Figure 9 A schematic diagram of the average connectivity coefficient of a water-driven reservoir generated by the generator.

[0029] Figure 10 A schematic diagram of the uncertainty distribution of the connectivity coefficient of a water-drive reservoir generated by the generator.

[0030] Figure 11 This is a graph showing the change of the comprehensive benefit objective function during the optimization process using a robust optimization model based on the particle swarm optimization algorithm in this embodiment.

[0031] Figure 12 This is a schematic diagram of the optimal robust injection and extraction scheme in this embodiment. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0033] This invention proposes a robust optimization method for water-drive reservoirs based on a data-driven physical model, such as... Figure 1 As shown, this method is applied to the dynamic injection-production optimization of a typical heterogeneous waterflood reservoir, specifically including the following steps: Step 1: Obtain the injection and production scheme for the water-driven reservoir within the target block. Divide the flow control grid according to the inter-well connectivity topology, construct a blocky capacitance-resistance physical model, determine the multi-level tensor iterative equations for multi-level series fluid transmission, and construct a priori dataset based on reservoir physical priors and statistical characteristics of water injection conditions. This includes the following sub-steps: Step 1.1: Obtain the injection and production scheme of a typical heterogeneous water-drive reservoir in this embodiment. The typical heterogeneous water-drive reservoir has 5 water injection wells (corresponding well numbers I1~I5) and 4 production wells (corresponding well numbers P1~P4). It has historical dynamic data of water injection and production over the past 3571 days. The goal is to formulate a robust injection and production scheme for the reservoir over the next 360 days based on the reservoir's historical dynamic data.

[0034] Based on the inter-well connectivity topology within the water-drive reservoir, the streamline trajectory from the injection well to the production well is divided. A series of cascaded flow control grids, namely the CRM-Block multi-level cascaded flow control grid, is used to characterize the multi-level delay effects of pressure signals and fluid transport in water-drive reservoirs. The grid is constructed including a connectivity coefficient matrix. and Group time constant matrix The physical model of a block capacitor and resistor, wherein, , , For the set of real numbers, This refers to the serial number of the flow control grid. In this embodiment, the total number of serial flow control grids is [number]. , .

[0035] Step 1.2: Based on the assumption of a pure water-driven steady-state oil reservoir and ignoring the primary oil recovery effect, determine the multi-level tensor iterative equation for multi-level series fluid transmission according to the discretized tensor forward transfer equation. This includes the flow response equations of the first-level grid and other grids in the series flow control grid, which are used to predict the total production matrix of the production well.

[0036] In this embodiment, in the multi-stage tensor iterative equation for multi-stage fluid transmission, the first-stage grid (i.e., the serial flow control grid number) in the serial flow control grid is used. When the value is 1, the flow response equation is determined as follows: ; In the formula, for The predicted output flow tensor of the first-level grid at time step; for The predicted output flow tensor of the first-level grid at time step; The product of Hadama; It is a natural exponential function; This represents the discretization time step. The time constant matrix of the first-level grid is used to characterize the physical hysteresis and attenuation effects when fluids pass through this level of grid in porous media; for The vector of actual water injection rates of each injection well in a water-driven reservoir at any given time; It is an all-one matrix with the same dimension as the time constant; This is the connectivity coefficient matrix, used to characterize the steady-state fluid distribution ratio from the injection well to the target production well; Used for characterization The initial effective fluid input that enters the first-level grid after connectivity allocation at any given moment.

[0037] For the other levels of the tandem flow control grid besides the first level grid (i.e., the tandem flow control grid sequence number) When the value is 2 or 3, the fluid input to the grid originates from the output of the previous level grid. The fluid undergoes exponential decay through an independent time constant matrix, and the flow response equation is determined as follows: ; In the formula, for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid.

[0038] Finally, the output of the last-level grid in the cascaded flow control grid is used as the total predicted production matrix of the production well.

[0039] Step 1.3: Based on the total predicted time series length of 3571 days for the water-drive reservoir and the total number of injection wells (5), the actual historical water injection sequence of the water-drive reservoir is extracted as a priori template. In the time dimension, abrupt change nodes are randomly set and a step scaling factor ranging from 0.75 to 1.35 is applied. In the numerical dimension, global and single-well independent scaling factors are introduced to nonlinearly perturb the priori template. Using a Dirichlet distribution, the total predicted time series length is randomly divided into multiple injection steps with a duration not less than the shortest stable period (i.e., 60 days). The injection rate of the first injection step ranges from 20% to 180%. Uniform sampling is performed within the injection stage, followed by random walk updates with a fluctuation factor applied to subsequent injection steps to simulate on-site allocation operations and generate a stepped water injection sequence matrix. .

[0040] Step 1.4: Construct physical parameter field labels, including a sparsed connectivity coefficient matrix. The set of time constant matrices that follow a log-normal distribution .

[0041] In this embodiment, a well-to-well connectivity sparsity constraint is introduced, along with the use of a Dirichlet distribution to generate the initial ratio, and an injection-end material conservation constraint is applied. This strictly constrains the sum of connectivity coefficients allocated from a single well to all production wells to within the range of 0.93 to 1.0, generating a connectivity coefficient matrix. .

[0042] The generated stepped water injection sequence matrix Substituting the physical parameter field labels into the multi-level tensor iterative equation, and superimposing a random Gaussian noise matrix positively correlated with its absolute magnitude onto the total liquid production matrix, yields the synthetic liquid production sequence with measurement errors. The synthetic product sequence Stepped water injection sequence matrix Prior samples are generated by combining the physical parameter field labels, and a prior dataset is constructed, which includes a stepped water injection sequence matrix. , synthetic liquid production sequence , connectivity coefficient matrix and the set of time constant matrices .

[0043] Step 2: Construct a generative adversarial network and input the prior samples from the prior dataset into the generative adversarial network to map and generate multiple sets of CRM-Block attribute parameter fields.

[0044] In this embodiment, the generative adversarial network includes a generator, a differentiable physical layer, and a discriminator.

[0045] The generator includes an input fusion layer, a temporal feature encoder, a self-attention module, and a multi-branch parameter decoder. The input fusion layer uses the actual water injection sequence of the water-drive reservoir. With actual liquid production sequence The injection-sampling time series is formed by concatenating the data along the feature channel dimension, and a random noise vector sampled from a standard normal distribution is introduced. Construct a multivariate input conditional feature matrix Conditional feature matrix The dimension is ,in, The scaling factor for the feature channel dimension is used to introduce a random noise vector as the core mechanism for driving the generative adversarial network to learn the conditional distribution of the ill-conditioned inverse problem and to achieve multi-solution sampling.

[0046] The temporal feature encoder contains five one-dimensional convolutional neural networks. The kernel size of each one-dimensional convolutional neural network is 4, the stride is 2, and the padding is 1. These networks are used to extract multi-scale local dynamic features and obtain deep convolutional feature maps.

[0047] Specifically, the forward propagation process of the one-dimensional convolutional neural network is as follows: ; In the formula, , , The first The hidden layer feature map, convolution kernel weight matrix, and bias term vector of a one-dimensional convolutional neural network; For the first Feature maps of hidden layers in a one-dimensional convolutional neural network; The LeakyReLU activation function has a fixed slope of 0.2 on the negative half-axis to avoid the gradient death problem. This is a one-dimensional convolution operation.

[0048] Considering that traditional convolutional maps are prone to losing the correlation between the first and last responses, the self-attention module in this embodiment uses a multi-head self-attention layer after the deep convolutional feature map to capture long sequence dependencies, and transforms the feature matrix into a query matrix through linear mapping. Key matrix Sum matrix The self-attention module is used to calculate the scaled dot product attention weights and output the context fusion features. .

[0049] Specifically, the context fusion feature for: ; In the formula, For the Softmax function; For querying the matrix; The key matrix; It is a value matrix; It is the transpose matrix; This is a scaling factor for the feature channel dimension, used to prevent the gradient of the Softmax function from vanishing due to excessively large dot product results.

[0050] The multi-branch parameter decoder is used to process context fusion features. The output generator outputs parameters and internally has a first branch and a second branch. The first branch obtains the connectivity coefficient matrix through the sigmoid activation function. , connectivity coefficient matrix The dimension is And the range of values ​​is The second branch obtains multiple sets of dimensions through the Softplus activation function. time constant matrix This forms a set of three-level time constant matrices. According to the connectivity coefficient matrix and multi-level time constant matrix set Output Generator Output Parameters .

[0051] Furthermore, the generator function in the generative adversarial network is set as follows: ; In the formula, For injection and extraction conditions, It is a random noise vector. Let be the connectivity coefficient matrix. It is a set of multi-level time constant matrices. This is a generator function.

[0052] The differentiable physical layer is used to substitute the generator output parameters into the multi-level tensor iterative equation for forward modeling, and backpropagate the predicted output flow tensor obtained from the forward modeling to the generator to achieve fully end-to-end training.

[0053] The discriminator is built on a least-squares adversarial architecture. First, the injected time series is input into three sequentially connected one-dimensional convolutional layers for downsampling, then input into an adaptive average pooling layer. After three rounds of one-dimensional convolutional downsampling and pooling compression, the data is flattened to obtain the temporal conditional feature vector. This is used to extract deterministic high-dimensional hidden features from the dynamic injection-production history of water-drive reservoirs within the target block; then the generator output parameters are... Flatten the input and map it to a fully connected layer as a static parameter feature vector. This is used to extract high-dimensional hidden features characterizing the topological structure of underground fluid connectivity; then a discriminator is used to compare the temporal conditional feature vectors. With static parameter eigenvectors The feature matching degree is used to determine whether the feature vector is true or false, and the temporal conditional feature vector is used for this purpose. With static parameter eigenvectors By splicing along the channel dimension, scalar scores are generated by outputting single-neuron linear layers without activation functions through a multilayer perceptron, and then mapped to generate multiple sets of CRM-Block attribute parameter fields.

[0054] Furthermore, the adversarial loss function of the discriminator in the generative adversarial network... Set to: ; In the formula, For mathematical expectation operations; For prior samples The distribution of physical parameter field label data conforms to real and geological a priori information; The linear confidence score is the score without the Sigmoid mapping. For random noise vectors It follows a standard normal distribution.

[0055] In this embodiment, the adversarial loss function of the discriminator In the middle, the first item The second term is used to encourage the discriminator to approach 1 in terms of linear confidence score for the real physical parameter field. This is used to encourage the discriminator to approach 0 in terms of linear confidence score for the generated physical parameter field.

[0056] Specifically, the total loss function of the generative adversarial network described in this embodiment Set to: ; in, ; ; ; In the formula, To counteract the hyperparameter weighting coefficients of the loss term, in this embodiment... ; To counteract the loss; The hyperparameter weighting coefficients for the physical response loss term are shown in this embodiment. ; The physical response loss term is used to constrain the generated physical parameter field to accurately fit the actual historical liquid production rate after being substituted into the differentiable physical layer for forward modeling. The hyperparameter weight coefficients of the parameter boundary constraint loss term are, in this embodiment... ; For parameter boundary constraint loss terms; In this embodiment, the hyperparameter weighting coefficients of the multi-solution diversity regularization loss term are... ; This is a loss term for multi-solution diversity regularization, used to penalize random noise vectors. Mode collapse behavior where changes in physical parameters, field parameters, and structural changes do not alter the generated physical parameters; For mathematical expectation operations; For random noise vectors It follows a standard normal distribution; The linear confidence score is the score without the Sigmoid mapping. For generator functions, , For injection and extraction conditions, It is a random noise vector. Let be the connectivity coefficient matrix. It is a set of multi-level time constant matrices. The time constant matrix; The linear confidence score is the score without the Sigmoid mapping. This represents the total number of time steps. This refers to the time step number; For production well serial numbers; Total number of producing wells; for Time of the first Predicted fluid production of a production well; for Time of the first The actual fluid production of a production well; It is a function for maximizing the value; It is an L1 norm; The preset disturbance response threshold is used to limit the minimum difference in the physical parameter fields generated by different random noise vectors. , All of these are generated physical parameter fields. , , , All of them are independent random noise vectors.

[0057] In this embodiment, the 5000 prior samples generated in step S1 are divided into training and validation sets in a 4:1 ratio. The batch size is set to 32, and the Adam optimizer is used with a fixed learning rate of 0.0002 for alternating adversarial training. To prevent gradient explosion in the physical layer during the initial training phase, a threshold of 5.0 is applied to the generator parameters during backpropagation as gradient clipping. After approximately 600 epochs of training iterations, the generator and discriminator reach Nash equilibrium. At this point, samples from the test set are extracted for validation, resulting in a comparison chart of the actual production volume of each production well and the predicted values ​​from the generative adversarial network, as shown below. Figures 2-5 As shown.

[0058] analyze Figures 2-5 Thus, in the inversion inference stage, the network weights of the generator are fixed after training, the historical injection and production sequence of the target water-driven reservoir is input, and 50 sets of random noise tensors independently sampled on the standard normal distribution are introduced. The generative adversarial network deterministically outputs 50 sets of physical parameter fields with physical consistency but structural differences within seconds of computation, realizing intelligent multi-solution sampling of uncertainty in high-dimensional nonlinear space.

[0059] A CRM-Block attribute parameter field is generated using a generative adversarial network generator based on the aforementioned 50 sets of physical parameter fields. This CRM-Block attribute parameter field includes connectivity coefficients and multi-level time constants. A heat map of the CRM-Block attribute parameter field is then obtained, as shown below. Figures 6-10 As shown.

[0060] Step 3: Generate a set of reservoir uncertainty scenarios based on the CRM-Block attribute parameter field, and set a feasible domain constraint space that satisfies the actual production and operation of the reservoir, using the decision variables in the reservoir injection and production scheme as the optimization objective. A robust optimization model was established to seek a robust injection-production scheme for the reservoir under the worst geological conditions, which includes the following sub-steps: Step 3.1: Set the total production forecast period for future reservoir optimization. For 360 days, time interval 30 days, controlled period There are 12, and they are arranged according to time intervals. The reservoir is discretely divided into 12 equal-length control periods. The injection volume of each water injection well in the reservoir during each future control period is used as a decision variable, resulting in a total of 60 decision variables. An injection volume decision is then constructed, yielding: ; In the formula, For the decision of injection volume, , For the set of real numbers, This represents the total number of injection wells. For the first Within the first control period, the first Injection rate of the injection well.

[0061] The generated multiple sets of CRM-Block attribute parameter fields are used as discrete reservoir uncertainty scenarios to construct a reservoir uncertainty scenario set. , ,in, This represents the sequence number of the reservoir uncertainty scenario. , This represents the total number of reservoir uncertainty scenarios in this embodiment. It is 50. The first in the set of reservoir uncertainty scenarios There are several reservoir uncertainty scenarios, and each reservoir uncertainty scenario includes a set of dimensions. The connectivity coefficient matrix and 3 sets of multi-level time constant matrices Each oil reservoir uncertainty scenario corresponds to a different geological dynamic response mode.

[0062] Step 3.2: For each reservoir uncertainty scenario in the reservoir uncertainty scenario set, obtain the injection volume decision, call the generative adversarial network, and use the differentiable physics layer in the generative adversarial network to perform forward modeling prediction to obtain the total production volume sequence of each production well in the reservoir within each specified control period. , .

[0063] Based on the total production volume sequence of each production well in the future reservoir during each specified control period, a comprehensive benefit objective function is established to characterize the water injection energy consumption and production revenue, with the injection volume of water injection wells as the decision variable and the total production volume of production wells as the objective. This function is used to achieve a quantitative evaluation of the reservoir's benefits from the injection end to the production end.

[0064] Specifically, the comprehensive benefit objective function is set as follows: ; In the formula, The objective function is the comprehensive benefit objective function; To control the time period sequence number; To control the total number of time periods; For production well serial numbers; Total number of producing wells; This refers to the serial number of the injection well; In this embodiment, the economic benefit weight of the unit liquid production is used. Yuan / cubic meter; In this embodiment, the energy consumption and operating cost per unit volume of water injection are weighted. Yuan / cubic meter; For the first Under the first reservoir uncertainty scenario, the first Predicted production rate of a production well; For the first Within the first control period, the first Injection rate of the injection well.

[0065] Step 3.3: Construct a feasible domain constraint space that satisfies the actual production and operation of the reservoir. .

[0066] In this embodiment, a feasible domain constraint space is constructed based on constraints such as single-well ultimate injection capacity, water injection capacity, production well fluid throughput, and on-site operational stability. Wherein, the single-well limit injection capacity constraint is the daily injection rate of each injection well. The water injection capacity constraint is set at the limit that the total injection volume of all injection wells must not exceed the maximum water supply capacity of the gathering and transportation station. The production well fluid throughput constraint is set to ensure that the predicted daily production fluid output of each production well does not exceed [a certain limit]. Meanwhile, to ensure a smooth transition when the optimized scheme is implemented in the oilfield and to avoid damage to surface equipment and underground rock mass caused by sudden changes in water injection network pressure, the stability constraint for on-site operation is set as follows: the adjustment range of the injection volume of each water injection well in the reservoir during adjacent control periods must be limited to the proportion boundary of safe fluctuation. Inside, Furthermore, the injection volume of each water injection well in the reservoir during the first control period must be smoothly connected with the actual water injection status at the end of the reservoir's production history.

[0067] Furthermore, by transforming robust decision-making theory into a mathematical optimization problem with a minimax game structure to determine robust reservoir injection and production schemes, a robust optimization model is constructed to find the robust injection and production scheme that maximizes the overall reservoir benefit objective function under the worst-case reservoir uncertainty scenario. .

[0068] In this embodiment, the robust optimization model makes decisions based on a preset injection volume. It traverses all reservoir uncertainty scenarios in the reservoir uncertainty scenario set, obtains the reservoir uncertainty scenario that minimizes the comprehensive benefit objective function, and takes it as the worst-case scenario. Then, it constrains the feasible region space. Robust injection volume decision for maximizing the overall benefit objective function in worst-case scenario for in-situ reservoir searching , ,in, The function is used to calculate the maximum value of the independent variable. It is a minimum value function.

[0069] Step 4: Utilize the particle swarm optimization algorithm within the feasible region constraint space that satisfies the actual production and operation of the reservoir. Solving the robust optimization model in the middle yields the optimal robust injection and acquisition scheme, which includes the following sub-steps: Step 4.1, Set the total number of particle swarm examples. and maximum number of generations In this embodiment , The decision on the dispensing quantity to be optimized is expanded and mapped to the particle positions in a high-dimensional search space. , Initialize the initial position and initial velocity of each particle in the particle swarm. The initial position of each particle corresponds to a specific candidate injection quantity decision scheme.

[0070] Step 4.2, utilizing particle swarm optimization in the feasible region constrained space The robust optimization model is solved through iterative calculation. In each iteration, each particle in the particle swarm sequentially traverses 50 reservoir uncertainty scenarios in the reservoir uncertainty scenario set, and calls the generative adversarial network to perform forward modeling to obtain the comprehensive benefit objective function value corresponding to each reservoir uncertainty scenario. The individual fitness of the particle in the current iteration step is determined based on the reservoir uncertainty scenario with the lowest comprehensive benefit objective function value.

[0071] Specifically, the formula for calculating the individual fitness of the particle is as follows: ; In the formula, For individual fitness evaluation function; For particle serial numbers; This refers to the iteration step number; For the first During the nth iteration The position of each particle; It is a minimum value function; The objective function is the comprehensive benefit objective function; For the first During the nth iteration Decision on the allocation of individual particles; This represents the sequence number of the reservoir uncertainty scenario. ; The first in the set of reservoir uncertainty scenarios Uncertainty scenario of oil reservoir.

[0072] Step 4.3: Use a robust optimization model to find the particle position that maximizes individual fitness, compare the current individual fitness of each particle with the historical individual fitness record, and update the historical best position of each particle in the particle swarm. And the global optimal position of the entire population. .

[0073] Then, determine the individual best position of each particle in the particle swarm. And the global optimal position of the entire population. As a guide, based on the inertia weights preset by the particle swarm optimization algorithm. Individual cognitive learning factors For social cognitive learning factors The particle swarm algorithm updates the velocity and position of each particle. In this embodiment, the inertia weight of the particle swarm algorithm is used during the particle velocity and position update process. Set to 0.8, Individual cognitive learning factor Set to 2.0, social cognitive learning factor Set to 2.0.

[0074] Step 4.4: Determine if the current iteration step meets the preset iteration termination condition. If the preset iteration termination condition is not met, return to step 4.2 to continue the iterative calculation; otherwise, set the currently determined global optimal position. It then reverse-maps and restores the injection volume decision, outputting the optimal robust injection and production scheme for all reservoir uncertainty scenarios in the reservoir uncertainty scenario set.

[0075] In this embodiment, the changes in the comprehensive benefit objective function during the optimization process using the robust optimization model based on the particle swarm optimization algorithm are as follows: Figure 11 As shown, after the particle swarm optimization algorithm terminates, the determined global optimal position is used... And its inverse mapping is restored to a dimension of The optimal robust injection and production scheme is obtained by deciding the injection volume, such as... Figure 12 As shown.

[0076] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A robust optimization method for water-drive reservoirs based on a data-driven physical model, characterized in that, Includes the following steps: Step 1: Obtain the injection and production scheme of the water-driven reservoir in the target block, divide the flow control grid according to the inter-well connectivity topology, construct the block capacitance and resistance physical model, determine the multi-level tensor iterative equation for multi-level series transmission of fluid, and construct the prior dataset based on reservoir physical priors and statistical characteristics of water injection conditions. Step 2: Construct a generative adversarial network and input the prior samples in the prior dataset into the generative adversarial network to map and generate multiple sets of CRM-Block attribute parameter fields; Step 3: Generate a set of reservoir uncertainty scenarios based on the CRM-Block attribute parameter field, and set a feasible domain constraint space that satisfies the actual production and operation of the reservoir, using the decision variables in the reservoir injection and production scheme as the optimization objective. Furthermore, a robust optimization model was established to seek robust injection and production schemes for reservoirs under the worst geological conditions. Step 4: Utilize the particle swarm optimization algorithm within the feasible region constraint space that satisfies the actual production and operation of the reservoir. Solve the robust optimization model to obtain the optimal robust injection and extraction scheme; The generative adversarial network includes a generator, a differentiable physical layer, and a discriminator; The generator includes an input fusion layer, a temporal feature encoder, a self-attention module, and a multi-branch parameter decoder. The input fusion layer uses the actual water injection sequence of the water-drive reservoir. With actual liquid production sequence The injection-sampling time series is formed by concatenating the data along the feature channel dimension, and a random noise vector sampled from a standard normal distribution is introduced. Construct a multivariate input conditional feature matrix The temporal feature encoder contains five one-dimensional convolutional neural networks to extract multi-scale local dynamic features and obtain deep convolutional feature maps. The self-attention module, by connecting a multi-head self-attention layer after the deep convolutional feature maps, is used to capture long sequence dependencies. The self-attention module calculates scaled dot product attention weights and outputs context fusion features. The multi-branch parameter decoder is used to process context fusion features. The output generator outputs parameters and internally has a first branch and a second branch. The first branch obtains the connectivity coefficient matrix through the sigmoid activation function. The second branch obtains the time constant matrix through the Softplus activation function. Form a set of multi-level time constant matrices According to the connectivity coefficient matrix and multi-level time constant matrix set Output Generator Output Parameters ; The differentiable physical layer is used to substitute the generator output parameters into the multi-level tensor iterative equation for forward modeling, and backpropagate the predicted output flow tensor obtained from the forward modeling to the generator to achieve fully end-to-end training. The discriminator is built on a least-squares adversarial architecture. First, the injected time series is input into three sequentially connected one-dimensional convolutional layers for downsampling, then input into an adaptive average pooling layer. After three rounds of one-dimensional convolutional downsampling and pooling compression, the data is flattened to obtain the temporal conditional feature vector. This is used to extract deterministic high-dimensional hidden features from the dynamic injection-production history of water-drive reservoirs within the target block; then the generator output parameters are... Flatten the input and map it to a fully connected layer as a static parameter feature vector. This is used to extract high-dimensional hidden features characterizing the topological structure of underground fluid connectivity; then a discriminator is used to compare the temporal conditional feature vectors. With static parameter eigenvectors The feature matching degree is used to determine whether the feature vector is true or false, and the temporal conditional feature vector is used for this purpose. With static parameter eigenvectors By splicing along the channel dimension, scalar scores are generated by outputting single-neuron linear layers without activation functions through a multilayer perceptron, and then mapped to generate multiple sets of CRM-Block attribute parameter fields.

2. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1.1: Based on the injection and production plan for the water-drive reservoir, obtain the total number of water injection wells for the water-drive reservoir. Total number of production wells The topological relationship between wells is defined by the streamline trajectory from the injection well to the production well. A series of flow control grids are used to characterize the multi-level delay effects of pressure signals and fluid transport in water-drive reservoirs, and a connectivity coefficient matrix is ​​constructed. and Group time constant matrix The physical model of a block capacitor and resistor, wherein, , , For the set of real numbers, This refers to the serial number of the flow control grid. , This represents the total number of serial flow control grids; Step 1.2: Based on the assumption of a pure water-driven steady-state oil reservoir and ignoring the primary oil recovery effect, determine the multi-stage tensor iterative equation for multi-stage series transmission of fluid according to the discretized tensor forward transfer equation, including the flow response equations of the first-stage grid and other-stage grids in the series flow control grid, which are used to predict the total production matrix of the production well. Step 1.3: Based on the injection and production scheme of the water-drive reservoir in the target block, extract the statistical characteristics of the water injection conditions as a priori template. In the time dimension, dynamically set the allocation mutation nodes based on a stochastic process. In the numerical dimension, introduce global and single-well independent scaling factors to nonlinearly perturb the priori template to generate a stepped water injection sequence matrix. ; Step 1.4: Construct physical parameter field labels, including a sparsed connectivity coefficient matrix. The set of time constant matrices that follow a log-normal distribution The generated stepped water injection sequence matrix Substituting the physical parameter field labels into the multi-level tensor iterative equation, and superimposing a random Gaussian noise matrix positively correlated with its absolute magnitude onto the total liquid production matrix, yields the synthetic liquid production sequence with measurement errors. The synthetic product sequence Stepped water injection sequence matrix Prior samples are generated by combining the physical parameter field labels with the prior samples to construct a prior dataset.

3. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 2, characterized in that, In the multi-stage tensor iterative equation for fluid multi-stage series transmission, the flow response equation is determined for the first-stage grid in the series flow control grid as follows: ； In the formula, for The predicted output flow tensor of the first-level grid at time step; for The predicted output flow tensor of the first-level grid at time step; The product of Hadama; It is a natural exponential function; This represents the discretization time step. The time constant matrix of the first-level grid; for The vector of actual water injection rates of each injection well in a water-driven reservoir at any given time; It is a matrix of all 1s; For the other levels of the cascaded flow control grid besides the first-level grid, the flow response equation is determined as follows: ； In the formula, for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid; for Time of the first The predicted output flow tensor of the grid.

4. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 1, characterized in that, The total loss function of the generative adversarial network Set to: ； in, ； ； ； In the formula, The hyperparameter weighting coefficients are used to counteract the loss term; To counteract the loss; These are the hyperparameter weighting coefficients for the physical response loss term; The physical response loss term is used to constrain the generated physical parameter field to accurately fit the actual historical liquid production rate after being substituted into the differentiable physical layer for forward modeling. These are the hyperparameter weight coefficients for the parameter boundary constraint loss term; For parameter boundary constraint loss terms; For the weighting coefficients of the hyperparameters of the multi-solution diversity regularization loss term; This is a loss term for multi-solution diversity regularization, used to penalize random noise vectors. Mode collapse behavior where changes in physical parameters, field parameters, and structural changes do not alter the generated physical parameters; For mathematical expectation operations; For random noise vectors It follows a standard normal distribution; The linear confidence score is the score without the Sigmoid mapping. For generator functions, , For injection and extraction conditions, It is a random noise vector. Let be the connectivity coefficient matrix. It is a set of multi-level time constant matrices. The time constant matrix; The linear confidence score is the score without the Sigmoid mapping. This represents the total number of time steps. This refers to the time step number; For production well serial numbers; Total number of producing wells; for Time of the first Predicted fluid production of a production well; for Time of the first The actual fluid production of a production well; It is a function for maximizing the value; It is an L1 norm; The preset disturbance response threshold is used to limit the minimum difference in the physical parameter fields generated by different random noise vectors. , All of these are generated physical parameter fields. , , , All of them are independent random noise vectors.

5. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Set the total production forecast period for future reservoir optimization. and according to time intervals Discretely divide it The injection volume of each water injection well in the reservoir during each control period of equal duration is used as a decision variable to construct an injection volume decision; The generated multiple sets of CRM-Block attribute parameter fields are used as discrete reservoir uncertainty scenarios to construct a reservoir uncertainty scenario set. , ,in, This represents the sequence number of the reservoir uncertainty scenario. , This represents the total number of uncertain scenarios in the oil reservoir. The first in the set of reservoir uncertainty scenarios There are several reservoir uncertainty scenarios, each including a set of connectivity coefficient matrices. as well as Group of multi-level time constant matrices Each reservoir uncertainty scenario corresponds to a different geological dynamic response mode; Step 3.2: For each reservoir uncertainty scenario in the reservoir uncertainty scenario set, obtain the injection volume decision, call the generative adversarial network, and use the differentiable physics layer in the generative adversarial network to perform forward modeling prediction to obtain the total production volume sequence of each production well in the reservoir within each specified control period. , ; Based on the sequence of total production volume of each production well in the reservoir during each specified control period, with the injection volume of water injection wells as the decision variable and the total production volume of production wells as the objective, a comprehensive benefit objective function is established to characterize the energy consumption of water injection and the revenue from production, so as to achieve a quantitative evaluation of the benefits of the reservoir from the injection end to the production end. Step 3.3: Construct a feasible domain constraint space that satisfies the actual production and operation of the reservoir. By transforming robust decision-making theory into a mathematical optimization problem with a minimax game structure to determine robust injection-production schemes for reservoirs, a robust optimization model is constructed to find robust injection-production schemes that maximize the overall reservoir benefit objective function under worst-case reservoir uncertainty scenarios. ; The robust optimization model, based on a preset injection volume decision, traverses all reservoir uncertainty scenarios in the reservoir uncertainty scenario set, identifies the scenario that minimizes the overall benefit objective function, and designates it as the worst-case scenario. Then, it performs optimization within the feasible region constraint space. Robust injection volume decision for maximizing the overall benefit objective function in worst-case scenario for in-situ reservoir searching , ,in, The function is used to calculate the maximum value of the independent variable. It is a minimum value function.

6. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 5, characterized in that, The comprehensive benefit objective function is set as follows: ； in, ； In the formula, The objective function is the comprehensive benefit objective function; For the decision of injection volume, , For the set of real numbers, This represents the total number of injection wells. For the first Within the first control period, the first Injection rate of the injection well; To control the time period sequence number; To control the total number of time periods; For production well serial numbers; Total number of producing wells; This refers to the serial number of the injection well; The economic benefit weight of the unit liquid production volume; Weighting of energy consumption and operating costs per unit volume of water injected; For the first Under the first reservoir uncertainty scenario, the first Predicted production rate of a production well; For the first Within the first control period, the first Injection rate of the injection well.

7. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 1, characterized in that, Based on constraints such as single-well ultimate injection capacity, water injection capacity, production well fluid throughput, and on-site operational stability, a feasible domain constraint space is constructed to meet the actual production and operation requirements of the reservoir. The single-well limit injection capacity constraint is limited by the rated power of the reservoir production equipment and the formation fracture pressure; the water injection capacity constraint is set to the total injection volume of all water injection wells in the reservoir within a single control period not exceeding the maximum water supply capacity of the gathering and transportation station; the production well fluid throughput constraint is set to the predicted production fluid volume not exceeding the maximum daily fluid discharge of the production well lifting equipment; the field operation stability constraint is set to the adjustment range of the injection volume of each water injection well in the reservoir during adjacent control periods must be limited within the proportional boundary of safe fluctuation, and the injection volume of each water injection well in the reservoir during the first control period must be smoothly connected with the actual water injection state at the end of the reservoir's production history.

8. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1, Set the total number of particle swarm examples. and maximum number of generations The decision on the dispensing quantity to be optimized is expanded and mapped to the particle positions in a high-dimensional search space. Initialize the initial position and initial velocity of each particle in the particle swarm; Step 4.2, utilizing particle swarm optimization in the feasible region constrained space In the process of iterative calculation to solve the robust optimization model, each particle in the particle swarm sequentially traverses each reservoir uncertainty scenario in the reservoir uncertainty scenario set, and calls the generative adversarial network to perform forward modeling to obtain the comprehensive benefit objective function value corresponding to each reservoir uncertainty scenario. The individual fitness of the particle in the current iteration step is determined based on the reservoir uncertainty scenario with the lowest comprehensive benefit objective function value. Step 4.3: Use a robust optimization model to find the particle position that maximizes individual fitness, compare the current individual fitness of each particle with the historical individual fitness record, and update the historical best position of each particle in the particle swarm. And the global optimal position of the entire population. ; Then, determine the individual best position of each particle in the particle swarm. And the global optimal position of the entire population. As a guide, update the velocity and position of each particle in the particle swarm; Step 4.4: Determine if the current iteration step meets the preset iteration termination condition. If the preset iteration termination condition is not met, return to step 4.2 to continue the iterative calculation; otherwise, set the currently determined global optimal position. It then reverse-maps and restores the injection volume decision, outputting the optimal robust injection and production scheme for all reservoir uncertainty scenarios in the reservoir uncertainty scenario set.

9. The robust optimization method for water-drive reservoirs based on a data-driven physical model according to claim 8, characterized in that, In step 4, the formula for calculating the individual fitness of the particle is: ； In the formula, For individual fitness evaluation function; For particle serial numbers; This refers to the iteration step number; For the first During the nth iteration The position of each particle; It is a minimum value function; The objective function is the comprehensive benefit objective function; For the first During the nth iteration Decision on the allocation of individual particles; This represents the sequence number of the reservoir uncertainty scenario. , This represents the total number of uncertain scenarios in the oil reservoir. The first in the set of reservoir uncertainty scenarios Uncertainty scenarios in oil reservoirs; The iteration termination condition is that the current iteration step reaches a preset maximum number of generations. Or the change in the fitness of the individual particle corresponding to the global optimal position is continuous. If the iteration rate is lower than the preset convergence threshold, it is determined that the preset iteration termination condition is met; otherwise, it is determined that the preset iteration termination condition is not met.

Images