A diffusion model-based method for inverting inter-well connectivity parameters and robust decision-making.

Abstract

This invention discloses a robust decision-making method for well connectivity parameter inversion based on a diffusion model, relating to the field of oil and gas field development technology. The invention first constructs a global parameter vector and performs logarithmic reparameterization to obtain standard input variables. These variables are then subjected to parameter baseline value calculation, hierarchical initial interval setting, dynamic envelope testing, and sensitivity hierarchical contraction. After adaptively optimizing the standard input variables to construct a standardized training dataset, a well connectivity parameter inversion model is constructed based on a graph-conditional dual-path diffusion inversion framework. The model is trained using the standardized training dataset, and uncertainty quantification is performed. A representative scenario set is constructed to obtain the reservoir development scheme to be optimized. A pre-defined reservoir dual-mode optimization model is solved within the representative scenario set to obtain uncertainty propagation quantification indicators and perform prior constraint dynamic feedback adjustments until the optimal reservoir development scheme is obtained, achieving accurate inversion of well connectivity parameters.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development technology, specifically to a method for inverting inter-well connectivity parameters and making robust decisions based on a diffusion model. Background Technology

[0002] The efficient development and long-term stable production of water-driven reservoirs depend on the precise formulation and dynamic optimization of injection and production schemes. The core of this lies in the detailed characterization of inter-well connectivity features and the efficient inversion of dynamic parameters. Traditional reservoir numerical simulation methods have rigorous physical mechanisms, but they suffer from problems such as cumbersome geological modeling, long historical fitting and prediction calculation cycles, and high computing costs, making it difficult to meet the needs of rapid on-site decision-making.

[0003] The GPSNet inter-well network proxy model can abstract the reservoir into a one-dimensional flow channel network between well points. It significantly simplifies calculations through equivalent two-dimensional mesh mapping, combining physical interpretability and high efficiency, and has become a mainstream technique for reservoir dynamic modeling and parameter inversion. However, in practical engineering applications, the GPSNet inter-well network proxy model still faces significant bottlenecks, primarily manifested in multiple solutions to parameter inversion, inefficient uncertainty quantification, and poor robustness of optimization decisions, severely limiting its field application. Regarding parameter prior constraints, traditional methods rely on manual empirical calculations, failing to distinguish between primary connectivity, ordinary connectivity, and weak connectivity, and employing a uniform constraint interval. This lacks physical rationality and a data-driven adaptive mechanism, easily introducing inversion errors from the outset. In high-dimensional parameter inversion, the GPSNet inter-well network proxy model has high-dimensional connectivity parameters, making deterministic inversion prone to getting trapped in local optima, outputting only a single solution, and completely ignoring parameter uncertainties. While random sampling methods such as MCMC and ensemble Kalman filtering can characterize uncertainty, the computational cost increases exponentially with dimensionality, iteration is time-consuming, and the ensemble size is difficult to select, failing to meet the requirements of real-time field decision-making. In terms of development scheme optimization, deterministic optimization is based on a single parameter solution and does not take into account geological uncertainties. In the middle and high water-cut stage, reservoirs are highly heterogeneous and have developed dominant channels. The injection water fingering and crossflow are obvious, resulting in poor robustness of the scheme and actual benefits far lower than the prediction. On the other hand, although traditional robust optimization takes risks into account, it requires massive simulation calculations, which are costly and do not establish a mechanism for quantifying uncertainty propagation, making it prone to over-optimization or under-optimization.

[0004] In recent years, generative AI technologies such as diffusion models and generative adversarial networks have provided new ideas for parameter inversion and uncertainty quantification. However, existing studies mostly treat well connectivity parameters as ordinary high-dimensional vectors, without designing a dedicated reverse denoising network based on the graph topology of GPSNet's "injection well-production well-well connectivity". They also fail to distinguish between the differences between the two types of physical properties, equivalent conductivity and equivalent connectivity volume, resulting in insufficient spatial consistency of the inversion results and easy compensatory bias in conductivity and volume. At the same time, existing methods do not form a closed-loop decision system, resulting in the optimization scheme being disconnected from uncertainty and unable to adapt to the actual scenario of GPSNet in dynamic modeling and development decision-making of water-drive reservoirs.

[0005] Therefore, in order to address the core pain points of current GPSNet parameter inversion and optimization decision-making, it is urgent to propose a diffusion model-based method for inverting and robustly deciding on inter-well connectivity parameters, so as to achieve accurate inversion of inter-well connectivity parameters and robust optimization of their development schemes. Summary of the Invention

[0006] This invention aims to solve the above-mentioned problems and proposes a method for inverting inter-well connectivity parameters and making robust decisions based on a diffusion model. By sequentially performing adaptive prior constraints, conditional diffusion inversion, posterior sample screening, robust optimization, and uncertainty feedback iteration, a closed-loop decision-making method for inter-well connectivity parameters is constructed. This method takes into account physical rationality, efficient uncertainty quantification, and low-cost robust decision-making, and achieves accurate inversion of injection and production parameters and robust optimization of development schemes for water-drive reservoirs, providing technical support for the refined and intelligent development of water-drive reservoirs.

[0007] The present invention adopts the following technical solution: A method for inverting inter-well connectivity parameters and making robust decisions based on a diffusion model includes the following steps: Step 1: Based on the actual well network layout of the target injection-production block, select the parameters to be inverted, construct a global parameter vector and perform logarithmic reparameterization to obtain the standard input variables; Step 2: Perform parameter baseline value calculation, hierarchical initial interval setting, dynamic envelope test and sensitivity hierarchical shrinkage on the standard input variable in sequence, adaptively optimize the value range of the standard input variable to obtain prior samples, and construct a standardized training dataset; Step 3: Based on the graph-conditional dual-path diffusion inversion framework that integrates production data, well connectivity topology, and physical properties, construct the well connectivity parameter inversion model and determine the multi-constraint loss function of the well connectivity parameter inversion model. Step 4: Train the well connectivity parameter inversion model using the standardized training dataset to obtain the trained well connectivity parameter inversion model, which is used to accurately learn the posterior distribution of well connectivity parameters. Use the trained well connectivity parameter inversion model to generate log parameter vector posterior samples and perform uncertainty quantification to construct a log parameter vector posterior sample set. Step 5: Select log-parameter vector posterior samples from the log-parameter vector posterior sample set to construct a representative scenario set to support subsequent optimization decisions; Step 6: Obtain the reservoir development scheme to be optimized, construct a dual-mode reservoir optimization model including a deterministic optimization model and a robust optimization model, solve the dual-mode reservoir optimization model in each representative scenario of the representative scenario set based on the intelligent optimization algorithm, obtain the uncertainty propagation quantification index of the dual-mode reservoir optimization model, and determine whether it exceeds the preset trigger threshold. If it exceeds the preset trigger threshold, trigger the prior constraint dynamic feedback adjustment mechanism; otherwise, output the optimized reservoir development scheme.

[0008] Preferably, step 1 includes the following sub-steps: Step 1.1: Based on the actual well network layout of the target injection-production block, obtain the total number of effective connections between wells within the target injection-production block. The equivalent conductivity and equivalent connectivity volume of each effective connection between wells are selected as parameters to be inverted. A global parameter vector is constructed to characterize the inter-well connectivity features of the target injection-production block, resulting in: ; In the formula, This is the global parameter vector; For the real number field; For the first Equivalent conductivity of effective connections between wells; For the first The equivalent connected volume of the effective connection between the strip wells; It is the transpose matrix; Step 1.2 introduces logarithmic reparameterization to eliminate training instability caused by the order-of-magnitude difference in the parameters to be inverted, and pre-sets the equivalent transmissivity benchmark value of the effective interconnection between wells based on the reservoir geology prior of the target injection-production block. and equivalent connected volume benchmark value By performing a logarithmic transformation on the global parameter vector and reconstructing the standard input variables, we obtain: ; In the formula, For standard input variables; For the first Log-reparameterized value of the equivalent conductivity of the effective connection between wells; For the first Logarithmic reparameterized value of the equivalent connected volume of the effective connection between wells.

[0009] Preferably, step 2 includes the following sub-steps: Step 2.1, for the effective connection between wells within the target injection-production block, determine the equivalent conductivity benchmark value based on the principles of seepage mechanics: ; In the formula, For the first The baseline value of the equivalent conductivity of the effective connection between wells; For the first Unit conversion and network equivalence correction coefficients for effective connections between wells; For the first Equivalent permeability index of effective connection between wells; For the first The equivalent flow cross-sectional area of ​​the effective connection between the shafts; For the first Equivalent fluid viscosity of the effective connection between wells; For the first The equivalent length of the effective connection between the strip wells; Based on reservoir volume characteristics, the equivalent connectivity volume benchmark value for effective connections between wells is determined as follows: ; In the formula, For the first The baseline value of the equivalent connectivity volume for effective connections between wells; For the first Volumetric equivalent correction factor for effective connection between shafts; For the first Equivalent porosity of effective connection between wells; For the first The equivalent thickness of the effective connection between the strip wells; Step 2.2: Based on the pre-defined prior constraint intervals according to the strength classification of inter-well connectivity, the effective connections between each well are divided into a primary connectivity candidate set, a normal connectivity candidate set, and a weak connectivity candidate set. The threshold values ​​for each type of connectivity candidate set in the prior constraint interval are set as follows: ; In the formula, For the first The lower bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells; For the first The upper bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells; The serial number of the effective connection between wells; Principal connected candidate set; The interval scaling factor of the logarithmic parameter vector of the effective connection between main wells; This is a candidate set for ordinary connected systems. This is the interval scaling factor for the logarithmic parameter vector of effective connections between ordinary wells; A candidate set of weakly connected networks; For the interval scaling factor of the logarithmic parameter vector of the effective connection between weak wells; ; Step 2.3: Optimize the prior constraint interval based on the dynamic envelope test, set the envelope indicator function and the prior envelope rate, and use Latin hypercube sampling to draw from the prior constraint interval. A set of prior samples are input into the GPSNet forward model to obtain simulated production vectors, which are then used as reservoir historical production observation data. Step 2.4: Perform parametric sensitivity analysis on each logarithmic parameter vector in the standard input variables to complete the graded shrinkage. Calculate the roughness sensitivity of each logarithmic parameter vector. Based on the roughness sensitivity of each logarithmic parameter vector in the standard input variables, classify each logarithmic parameter vector into high sensitivity, medium sensitivity, and low sensitivity. Determine the differential shrinkage factor based on the sensitivity of each logarithmic parameter vector. Determine the shrunken prior interval, obtain the filtered prior parameter samples, and construct a standardized training dataset. The formula for calculating roughness sensitivity is: ; In the formula, For standard input variables The Middle The coarseness sensitivity of a logarithmic parameter vector; This refers to the prior sample group number; The first in the standard input variables Group of prior samples; The perturbation scale; For the first Unit vectors; For reservoir history well control vectors; For 2-norm operations; This is the GPSNet forward mapping model.

[0010] Preferably, step 3 includes the following sub-steps: Step 3.1: Using historical production observation data of the reservoir as a condition and learning the posterior distribution of the logarithmic parameter vector as the inversion target, obtain the actual observation vector of the reservoir, resulting in: ; In the formula, This represents the actual observed vector of the reservoir. For a moment, , This represents the total reservoir production time. For production well serial number, , Total number of producing wells; This refers to the serial number of the water injection well. , This represents the total number of injection wells. for Time of the first Measured oil production of a production well; for Time of the first Measured bottomhole flowing pressure of a production well; for Time of the first Measured injection pressure of the injection well; It is the transpose matrix; Step 3.2: Set up a Markov chain noise-adding mechanism, construct a diffusion model to simulate the forward noise-adding diffusion process, and generate a noisy logarithmic parameter vector, resulting in: ; In the formula, This represents the conditional probability distribution of the forward noise diffusion process; This is the numbering of the diffusion time step. , This represents the total number of diffusion time steps. For the first The logarithmic parameter vector after adding noise; For the first The logarithmic parameter vector after adding noise; It follows a multivariate normal distribution; For the first The noise schedule factor of the step; It is the identity matrix; Step 3.3: Based on the graph conditional dual-path diffusion inversion framework, construct the well connectivity parameter inversion model, including a conditional reverse denoising network that fuses production data encoders, a multi-condition fusion layer, and a conductivity-volume physics decoupled dual-path denoising network; Based on the reservoir historical production observation data, output conditional feature vectors are generated, and then a graph conditional denoising network based on the inter-well connectivity topology is constructed to generate an inter-well connectivity topology map. After multi-condition fusion of the noisy parameter vector, conditional feature vector, and inter-well connectivity topology map, the data is input into the conductivity-volume physical decoupled dual-path denoising network to invert and obtain the posterior samples of inter-well connectivity parameters, including the equivalent conductivity and equivalent connectivity volume of each effective connection between wells. Step 3.4: Construct the multi-constraint loss function for the well connectivity parameter inversion model.

[0011] Preferably, the production data encoder is provided with four one-dimensional convolutional layers and a Transformer self-attention layer, which are used to extract conditional feature vectors from historical reservoir production observation data, capture long-range temporal dependencies and core features in historical reservoir production observation data, input historical reservoir production observation data into the production data encoder, and use the production data encoder to process and output conditional feature vectors. The graph conditional denoising network is set as an inverse denoising network, using GAT as the backbone, to obtain the well connectivity topology graph, resulting in: ; in, ; ; In the formula, This is a topology diagram showing the connectivity between wells; For the set of well nodes; For the set of water injection well nodes; For the set of production well nodes; This is the set of effective connecting edges between wells; The serial number of the effective connection between wells; The first Logarithmic parameter vector after step-by-step noise addition Based on physical properties, they are divided into logarithmic subvectors of conductivity. and connected volume logarithmic subvector and will the After adding noise, the noisy conductivity component and noisy volume component of the effective connection between each well are used as edge features. Combined with the node features of each injection well node and each production well node, the standardized encoding is obtained as follows: ; In the formula, For the first The initial edge feature vector of the effective connection between the strips; , The first After adding noise, the first step Logarithmic component of noisy conductivity and logarithmic component of noisy connectivity volume in the effective connection between wells; For nodes The initial node feature vector; Encode and map node features; For nodes The historical production statistics feature vector of the corresponding well; To produce the conditional feature vector output by the data encoder, ,in, To produce data encoders, To generate network parameters for the data encoder, This represents the actual observed vector of the reservoir. For diffusion time step Location encoding.

[0012] Preferably, reverse denoising is performed by using a conductivity-volume physical decoupled dual-path denoising network to execute node-edge joint message passing. The reverse denoising process is as follows: ; In the formula, This represents the conditional probability distribution for the reverse denoising process; It is the mean function of the inverse denoising distribution; For the first The variance coefficient of the inverse denoising distribution after step-by-step noise addition; By aggregating adjacent edges and neighboring node information, the edge feature vectors of effective connections between each well and the feature vectors of each node are updated to obtain the updated edge feature vectors of effective connections between each well and the feature vectors of each node. The feature vector update process is as follows: ; ; In the formula, The layer numbering is used for the graph network. , All are node serial numbers; , They are nodes and nodes In the Node representation of a layer in a layered graph network; It is a non-linear activation function; For nodes Adjacent connection sets; , The first Layered graph network layer, first The first layer in a layered graph network Graph attention weights for effective connections between wells; The message transformation matrix; Update the mapping for edge features; , The first In a layered graph network, the layer located at the first Effective connection between the two ends of the injection well node between the strip wells Production well nodes The node representation; The conductivity-volume physical decoupled dual-path denoising network includes a conductivity network configured in parallel. and volume network The intermediate hidden representation of the conductivity network is obtained. Intermediate hidden representation of volumetric networks , , ,in, For the set of network parameters of a conductivity network, This is the set of network parameters for the volumetric network. In conductivity network and volume network A cross-attention fusion layer is introduced into the intermediate layer to allow conductivity information and volume information to be mutually corrected, which is used to express the synergistic effect of conductivity and reservoir supply capacity in seepage mechanics, resulting in: ; ; In the formula, This is the hidden representation of the conductance network after cross-attention fusion; This represents the hidden representation of the volumetric network after cross-attention fusion; Calculate operators for attention; , , , , , All are learnable linear mapping matrices; For conductivity networks Predicted conductance noise and volumetric networks Predicted volume noise The predicted noise is obtained by splicing. , ,in, This is a vector concatenation operation; The predicted noise is obtained after processing using a conductivity-volume physical decoupled dual-path denoising network. , Furthermore, the predicted noise is used as an explicit condition for the inverse denoising distribution.

[0013] Preferably, the multi-constraint loss function is set as follows: ; in, ; ; ; ; In the formula, For the multi-constraint loss function value; Predict the noise loss for the conductivity network; Predict the noise loss for volumetric networks; These are the weighting coefficients for the interval constraint loss; For interval-constrained soft loss with logarithmic parameters; The weighting coefficients for the physical coupling consistency loss; This represents a loss of consistency due to physical coupling. For expectation calculation; The true logarithmic parameter vector without noise; This is the number of the diffusion time step; This represents the true Gaussian noise component corresponding to the logarithmic subvector of conductivity. Transitivity noise predicted by the transitivity network; Number the time steps of the diffusion process; The conditional feature vector output by the production data encoder; The true Gaussian noise component corresponding to the logarithmic subvector of the connected volume; Volume noise predicted by the volume network; This refers to the 2-norm square operation. The index of the logarithmic parameter vector; This represents the total number of effective connections between wells. It is a function for maximizing the value; For the first A logarithmic parameter vector; For the first The upper bound of the final constraint interval for the logarithmic parameter vector; For the first The lower bound of the final constraint interval for the logarithmic parameter vector; The normalized weight matrix for the observations; Forward mapping model of GPSNet; For the first The noisy parameters after adding noise and the noiseless logarithmic parameter estimates obtained by restoring the predicted noise.

[0014] Preferably, step 4 includes the following sub-steps: Step 4.1: Train the well connectivity parameter inversion model using a standardized training dataset. During training, implement posterior iterative shrinkage to improve the inversion accuracy of the well connectivity parameter inversion model. After the well connectivity parameter inversion model completes each round of training and inversion, it generates a dataset containing... The posterior sample set of the posterior samples of the inter-well connectivity parameters is used to take the logarithmic parameter vectors in the posterior sample set. The parameter interval is iteratively shrunken by the quantile interval to obtain the trained well connectivity parameter inversion model, and the updated quantile interval is used as the parameter constraint. The iterative shrinkage process of the parameter interval is as follows: ; In the formula, For the first The upper bound of the final constraint interval for the logarithmic parameter vector; For the first The lower bound of the final constraint interval for the logarithmic parameter vector; for Quantile operations; For the posterior sample of all well connectivity parameters, the first... The set of possible values ​​for a logarithmic parameter vector; This is the quantile threshold; Step 4.2 employs a differential implicit accelerated sampling algorithm, utilizing the well connectivity parameter inversion model to accelerate the generation of posterior samples for well connectivity parameters. It initializes the noisy logarithmic parameter vector based on a standard normal distribution, and simultaneously inputs the current diffusion time step, the noisy parameter vector, the conditional feature vector, and the well connectivity topology map at each inverse diffusion step. The trained well connectivity parameter inversion model is then used to predict noise, and denoising is performed progressively according to the inverse diffusion time step. This is combined with the DDIM accelerated sampling algorithm to reduce the number of sampling steps and improve sampling efficiency, generating... The standard input variables are set and uncertainty quantification analysis is performed to construct a logarithmic parameter vector posterior sample set.

[0015] Preferably, in step 5, logarithmic parameter vector posterior samples are screened from the logarithmic parameter vector posterior sample set. A mismatch function is used to quantify the fit between the logarithmic parameter vector posterior samples and historical reservoir observation data. Invalid logarithmic parameter vector posterior samples are then removed from the logarithmic parameter vector posterior sample set to obtain... Group effective posterior sample set ; The mismatch function is set as follows: ; In the formula, For the first The reservoir history fitting mismatch function value of the group of effective posterior sample sets; These are the weighting coefficients for the yield fit; For the first The simulated output vector obtained from the effective posterior sample set of the group through the GPSNet forward mapping model; This represents the field-measured vector of reservoir production. This refers to the 2-norm square operation. For the first The simulated pressure vector is obtained from the effective posterior sample set through the GPSNet forward mapping model; This is the field-measured vector of pressure; Then construct a mean parameter model, and use the mean parameter model to analyze each effective posterior sample set. Perform mean calculation to determine the logarithmic parameter mean vector. The calculation formula is: ; In the formula, This is a vector of logarithmic parameter means, including the mean of the equivalent conductivity of the effective connections between each well. and equivalent connected volume ; For the effective logarithmic parameter vector posterior samples in the effective posterior sample set; Then, by inverse logarithmic transformation, the mean equivalent conductivity and mean equivalent connectivity volume of the effective connections between each well are obtained, resulting in: ; In the formula, For the first Mean equivalent conductivity of effective connections between wells; For the first The baseline value of the equivalent conductivity of the effective connection between wells; For the first Mean equivalent conductivity of effective connections between wells; For the first Mean equivalent connectivity volume of effective connections between wells; For the first The baseline value of the equivalent connectivity volume for effective connections between wells; For the first The equivalent connected volume of the effective connection between the strip wells; It is the Euler number; Finally, the K-means hierarchical clustering algorithm was used to analyze each effective posterior sample set. Cluster analysis was performed by using the effective posterior sample set The effective logarithmic parameter vector in the posterior sample is divided into _ ... For each category, select the posterior sample of the effective log parameter vector closest to the cluster center to construct a representative scene set. , , Representative scene set The Middle The posterior sample of the effective log-parameter vector that is closest to the cluster center in each category.

[0016] Preferably, step 6 includes the following steps: Step 6.1: Obtain the reservoir development scheme to be optimized and determine the decision variables of the reservoir development scheme to be optimized. And in conjunction with the engineering physical feasible region constraints that determine the decision variables. A net present value objective function is constructed using the reservoir net present value as the development objective. ; The net present value objective function Set to: ; In the formula, Representative scene set The Middle The posterior sample of the effective log-parameter vector closest to the cluster center in each category; This refers to the time step number; This represents the total number of time steps. The profit coefficient per unit of oil production; for Time of the first Simulated oil production in representative scenarios; Cost coefficient per unit volume of water injected; for Time of the first Simulated water injection volume for a representative scenario; Cost coefficient per unit of production operation; for Time of the first Simulated production operation volume for a representative scenario; The annual discount rate; For time step The corresponding absolute time; For time scale conversion factors; Step 6.2: Construct a dual-mode optimization model for the reservoir, including a deterministic optimization model and a robust optimization model, wherein the deterministic optimization model is set as follows: , It is a function with maximum value. The mean vector of logarithmic parameters; the robust optimization model is set as follows: ,in, , , Decision variables Expected returns in representative scenarios; The serial number represents the scene. The total number of representative scenarios; Decision variables Variance of returns in representative scenarios; Risk aversion coefficient; Step 6.3: In all representative scenarios of the representative scenario set, an intelligent optimization algorithm is used to optimize the decision variables within the engineering-physical feasible region constraints. Solve deterministic and robust optimization models in the algorithm, and obtain the uncertainty propagation quantification indexes for the deterministic and robust optimization models, including the control drift index. and return dispersion index , , ,in, For the first The optimal decision variable matrix obtained by individual optimization in a representative scenario; This is the mean matrix of the optimal decision matrices for all representative scenarios. ; This refers to the 2-norm square operation. Step 6.4: Based on the uncertainty propagation quantification index of the deterministic optimization model and the robust optimization model, using the uncertainty propagation quantification index of the deterministic optimization model as the standard value, calculate the difference in control drift index and the difference in revenue dispersion index between the robust optimization model and the deterministic optimization model, and compare them with the preset trigger threshold. If the preset trigger threshold is exceeded, the prior constraint dynamic feedback adjustment mechanism is triggered, using the uncertainty propagation quantification index of the robust optimization model as the feedback signal. The prior constraint interval, sensitivity interval scaling factor, and prior envelope rate are adjusted according to the uncertainty propagation quantification index of the robust optimization model. Steps 1 to 6 are re-executed until the prior constraint dynamic feedback adjustment mechanism is no longer triggered, and the optimized reservoir development scheme is output. If the preset trigger threshold is not exceeded, the prior constraint dynamic feedback adjustment mechanism is not triggered, and the optimized reservoir development scheme is directly output.

[0017] The present invention has the following beneficial effects: (1) This invention proposes a method for inverting and robustly deciding on well connectivity parameters based on a diffusion model. This method overcomes the subjective limitations of traditional manual parameter range setting by constructing a benchmark value, envelope testing, sensitivity classification and posterior iteration, and realizes physically reasonable and data-driven adaptive adjustment of parameter range, thus alleviating the problem of multiple solutions in well connectivity parameter inversion from the source.

[0018] (2) This invention proposes a diffusion model-based method for inverting inter-well connectivity parameters and robust decision-making. Based on the GPSNet conditional diffusion inversion model, an inter-well connectivity parameter inversion model with physical constraints is constructed. By reparameterizing the logarithm, using interval constraint soft loss and posterior iterative contraction, high-precision inversion of inter-well connectivity parameters and second-level uncertainty sampling are achieved. Compared with the traditional deterministic inversion method, which provides complete parameter distribution information, the loss function of the model constructed in this invention integrates multi-dimensional constraints, effectively ensuring the physical consistency of the inversion parameters. By establishing a full-process real closed-loop decision-making system of "inversion-screening-clustering-optimization-feedback iteration", and by using representative scenario sets to balance uncertainty coverage and computational cost in the inter-well connectivity parameter inversion process, a reservoir dual-mode optimization model including a deterministic optimization model and a robust optimization model is constructed. The robust optimization model is used to maximize expected returns while minimizing risk fluctuations.

[0019] (3) This invention proposes a robust decision-making method for inversion of inter-well connectivity parameters based on a diffusion model. It innovatively introduces an uncertainty propagation quantification index and constructs a closed-loop feedback mechanism, scientifically judges the application scenarios of robust optimization, and transmits the uncertainty at the decision end back to the prior constraint link for iterative optimization, avoiding over-optimization or under-optimization. It can provide flexible and adaptable reservoir development schemes for different complex reservoirs. The uncertainty propagation quantification index scientifically quantifies the impact of uncertainty from the dual dimensions of reservoir development scheme decision variables and economic benefits, making the inversion results of inter-well connectivity parameters more in line with the actual situation, realizing accurate inversion of inter-well connectivity parameters, and providing technical support for the refined and intelligent development of complex reservoirs. Attached Figure Description

[0020] Figure 1 This is a flowchart of a well connectivity parameter inversion and robust decision-making method based on a diffusion model according to the present invention.

[0021] Figure 2 This is a schematic diagram of the interconnected topology between wells within the water-driven reservoir block in this embodiment.

[0022] Figure 3 This is a schematic diagram of the structure of the well connectivity parameter inversion model.

[0023] Figure 4 The figure shows the uncertainty quantification results of the equivalent transmission rate inversion. In the figure, (a) is the uncertainty quantification result of the equivalent transmission rate inversion of the standard input variables in groups 1 to 10, and (b) is the uncertainty quantification result of the equivalent transmission rate inversion of the standard input variables in groups 11 to 20.

[0024] Figure 5The figure shows the uncertainty quantification results of the equivalent connected volume inversion. In the figure, (a) is the uncertainty quantification result of the equivalent connected volume inversion of the standard input variables for groups 1 to 10, and (b) is the uncertainty quantification result of the equivalent connected volume inversion of the standard input variables for groups 11 to 20.

[0025] Figure 6 The diagram shows the fitting results of reservoir production data for production well P1. In the diagram, (a) is the fitting curve of daily oil production of production well P1, (b) is the fitting curve of formation pressure of production well P1, and (c) is the fitting curve of water cut of production well P1.

[0026] Figure 7 The diagram shows the fitting results of reservoir production data for production well P2. In the diagram, (a) is the fitting curve of daily oil production of production well P2, (b) is the fitting curve of formation pressure of production well P2, and (c) is the fitting curve of water cut of production well P2.

[0027] Figure 8 This is a schematic diagram of the convergence curve during the optimization process of the particle swarm optimization algorithm in this embodiment.

[0028] Figure 9 This is a schematic diagram illustrating the control drift indices of the robust optimization model and the deterministic optimization model.

[0029] Figure 10 This is a schematic diagram illustrating the revenue dispersion index for robust optimization models and deterministic optimization models. Detailed Implementation

[0030] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0031] This invention proposes a method for inverting inter-well connectivity parameters and making robust decisions based on a diffusion model, such as... Figure 1 As shown, this method was applied to a heterogeneous water-drive reservoir block with medium to high water cut. This water-drive reservoir block contains 6 injection wells (well numbers I1-I6) and 8 production wells (well numbers P1-P8). Based on the inter-well seepage patterns, 20 effective connections between wells were identified, and historical production dynamics data for the past 800 days were obtained, including reservoir production, pressure, and water injection volume. Using the method of this invention, inter-well connectivity parameters were inverted based on historical reservoir dynamics, and a robust injection-production development plan for the water-drive reservoir was formulated for the next 120 days. The specific steps include: Step 1: Based on the actual well network layout of the target injection-production block, select the parameters to be inverted, construct a global parameter vector, and perform logarithmic reparameterization to obtain the standard input variables. This includes the following sub-steps: Step 1.1: Based on the actual well network layout of the target injection-production block, obtain the total number of effective connections between wells within the target injection-production block. In this embodiment, fixed parameter configuration for the water-drive reservoir block is completed based on the reservoir numerical simulation data file. The grid system for the water-drive reservoir block adopts a Cartesian grid, with 50, 40, and 1 grid cells for the X, Y, and Z axes, respectively. The length and width of the planar grid are set to 20m. The initial water saturation of the water-drive reservoir block is set to 0.2, the effective thickness to 2.0m, the horizontal permeability to 120mD, and the porosity to 0.2. The crude oil viscosity in the formation fluid parameters of the water-drive reservoir block is 5.2. The well spacing constraint for the connection between the 6 water injection wells and the 8 production wells is 350 meters. Based on the inter-well seepage law, 20 effective inter-well connections are constructed within the water-drive reservoir block, such as... Figure 2 As shown.

[0032] In this embodiment, the equivalent conductivity and equivalent connectivity volume of each effective connection between wells are used as parameters to be inverted. By inverting and anchoring the seepage capacity and reservoir carrying capacity of the water-drive reservoir block, the key characteristics of fluid migration between wells are accurately characterized.

[0033] Furthermore, a global parameter vector is constructed to characterize the inter-well connectivity features of the target injection-production block. The global parameter vector has a dimension of 120 and is expressed as follows: ; In the formula, This is the global parameter vector, i.e., the global parameter vector to be inverted for GPSNet well connectivity; It is the field of real numbers, dimensionless; For the first The equivalent conductivity of the effective connection between wells, used to characterize seepage capacity, is expressed in units of... ; For the first The equivalent connectivity volume of the effective connections between wells is used to characterize the reservoir's carrying capacity, and the unit is 1. ; This is the transpose of the matrix.

[0034] Step 1.2 introduces logarithmic reparameterization to eliminate training instability caused by the order-of-magnitude difference in the parameters to be inverted, and pre-sets the equivalent transmissivity benchmark value of the effective interconnection between wells based on the reservoir geology prior of the target injection-production block. and equivalent connected volume benchmark value By performing a logarithmic transformation on the global parameter vector, we can reconstruct the standard input variables that the diffusion model can directly learn, resulting in: ; in, ; ; In the formula, Standard input variable, dimensionless; For the first The logarithmic reparameterized value of the equivalent conductivity of the effective connection between wells, dimensionless; For the first The logarithmic multiparameterized value of the equivalent connected volume of the effective connection between wells, dimensionless; It is the natural logarithm function.

[0035] Step 2: Construct an adaptive prior constraint system. This involves sequentially calculating baseline parameter values, setting tiered initial intervals, performing dynamic envelope testing, and sensitivity-tiered contraction on the standard input variables. This adaptively optimizes the range of values ​​for the standard input variables to obtain prior samples and constructs a standardized training dataset. This includes the following sub-steps: Step 2.1: Construct a parameter benchmark value calculation system with clear physical meaning. For the effective connection between wells within the target injection-production block, determine the equivalent conductivity benchmark value based on the principles of seepage mechanics: ; In the formula, For the first The equivalent conductivity benchmark value for effective connection between wells is dimensionless. For the first Unit conversion and network equivalence correction coefficients for effective connections between wells, dimensionless; For the first The equivalent permeability index of the effective connection between wells, in units of ; For the first The equivalent flow cross-sectional area of ​​the effective connection between the wells, in units of ; For the first Equivalent fluid viscosity of the effective connection between wells, in units of ; For the first The equivalent length of the effective connection between the strips is expressed in units of... .

[0036] Based on reservoir volume characteristics, the equivalent connectivity volume benchmark value for effective connections between wells is determined as follows: ; In the formula, For the first The equivalent connected volume benchmark value of the effective connection between wells is dimensionless. For the first The volume equivalent correction factor for the effective connection between wells is dimensionless. For the first The equivalent porosity of the effective connection between wells is dimensionless. For the first The equivalent thickness of the effective connection between the strips is expressed in units of... .

[0037] Step 2.2: Based on the pre-defined prior constraint intervals according to the strength classification of inter-well connectivity, the effective connections between each well are divided into a primary connectivity candidate set, a normal connectivity candidate set, and a weak connectivity candidate set. The threshold values ​​for each type of connectivity candidate set in the prior constraint interval are set as follows: ; In the formula, For the first The lower bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells is dimensionless. For the first The upper bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells is dimensionless. The serial number of the effective connection between wells; Principal connected candidate set; The interval scaling factor of the logarithmic parameter vector of the effective connection between main wells, dimensionless; This is a candidate set for ordinary connected systems. , is the interval scaling factor for the logarithmic parameter vector of effective connections between ordinary wells, and is dimensionless; A candidate set of weakly connected networks; , is the interval scaling factor for the logarithmic parameter vector of the effective connection between weak wells, and is dimensionless.

[0038] In this embodiment, based on well spacing, dynamic response, and geological steering, the 20 effective well connections are divided into 6 main well connections, 8 ordinary well connections, and 6 weak well connections. The interval scaling factor of the logarithmic parameter vector of the main well connections is then used. Set to 0.8, the interval scaling factor for the effective logarithmic parameter vector between ordinary wells. Set to 1.0, the interval scaling factor for the logarithmic parameter vector of the effective connection between weak wells. Set it to 1.5.

[0039] Step 2.3: Optimize the prior constraint interval based on dynamic envelope test, set the envelope indicator function and prior envelope rate, and use Latin hypercube sampling to extract 500 sets of prior samples from the prior constraint interval. Input the prior samples into the GPSNet forward model to obtain the simulated production vector and use it as the reservoir historical production observation data, thereby providing data support for subsequent interval tests.

[0040] Specifically, the envelope indication function is set as follows: ; In the formula, For a moment, , This represents the total reservoir production time. The serial number of the reservoir production observation variable; for Time of the first The first group of prior samples The envelope indicator function of each reservoir production observation variable is dimensionless. for Time of the first Dimensionless field measured values ​​of reservoir production observation variables; for All prior samples at time n The simulated minimum value of each reservoir production observation variable, dimensionless; for All prior samples at time n The simulated maximum value of each reservoir production observation variable, dimensionless.

[0041] The prior envelope rate is set as follows: ; In the formula, The envelope of the prior constraint interval is used to represent the proportion of measured values ​​falling within the simulated value interval; it is dimensionless. The total dimension of the reservoir production observation variables is dimensionless. For envelope indicator function, when If the prior constraint interval is expanded, it is expanded; otherwise, it is contracted, so that the prior constraint interval can accurately cover the historical production observation data of the reservoir.

[0042] Step 2.4: Perform parametric sensitivity analysis on each logarithmic parameter vector in the standard input variables to complete the graded shrinkage. Calculate the roughness sensitivity of each logarithmic parameter vector. Based on the roughness sensitivity of each logarithmic parameter vector in the standard input variables, classify each logarithmic parameter vector into high sensitivity, medium sensitivity, and low sensitivity. Determine the differential shrinkage factor based on the sensitivity of each logarithmic parameter vector. The shrinkage prior interval is determined, the filtered prior parameter samples are obtained, and a standardized training dataset is constructed.

[0043] Specifically, the roughness sensitivity calculation formula is as follows: ; In the formula, For standard input variables The Middle The coarseness sensitivity of a logarithmic parameter vector is used to characterize the degree of influence of the parameters on the observed values, and is dimensionless. This refers to the prior sample group number; The first in the standard input variables Group of prior samples; The perturbation scale is dimensionless; For the first A unit vector, dimensionless; The reservoir's historical well control vector is dimensionless. This is a 2-norm operation used to characterize the magnitude of a vector; it is dimensionless. This is a GPSNet forward mapping model used to map logarithmic parameter vectors and reservoir historical well control vectors to reservoir production responses.

[0044] In this embodiment, hierarchical shrinkage is performed based on parameter sensitivity analysis. The coarse sensitivity of each logarithmic parameter vector is calculated, and the 40 sets of logarithmic parameter vectors are divided into 12 sets of highly sensitive logarithmic parameter vectors, 16 sets of moderately sensitive logarithmic parameter vectors, and 12 sets of low-sensitivity logarithmic parameter vectors. By dividing the logarithmic parameter vectors into three types—highly sensitive, moderately sensitive, and low-sensitivity—shrinkage factors are set for each type. Specifically, the shrinkage factor for highly sensitive logarithmic parameter vectors is set to 0.4, the shrinkage factor for moderately sensitive logarithmic parameter vectors is set to 0.7, and the shrinkage factor for low-sensitivity logarithmic parameter vectors is set to a fixed baseline value. This further compresses the solution space and improves the inversion stability and inversion calculation efficiency.

[0045] Step 3: Based on the graph-conditional bipath diffusion inversion framework that integrates production data, well connectivity topology, and physical properties, construct the well connectivity parameter inversion model and determine the multi-constraint loss function of the well connectivity parameter inversion model, including the following sub-steps: Step 3.1 uses historical reservoir production observation data as a condition and learns the posterior distribution of the logarithmic parameter vector as the inversion target, ensuring that the inversion revolves around the actual reservoir production data, effectively improving the fit between the inversion results and the actual production characteristics. In this embodiment, the actual reservoir observation vector is obtained based on 800 days of historical reservoir production data and used as the conditional signal for the diffusion model. This vector includes the oil production, bottomhole flowing pressure, and water cut of 8 production wells, with a total dimension of 24×800. It is used to drive the learning of the posterior distribution of the logarithmic parameter vector, while simultaneously achieving a quantitative representation of parameter uncertainty, effectively solving the problem that traditional deterministic inversion methods cannot reflect parameter distribution.

[0046] In this embodiment, the actual observation vector of the reservoir is obtained based on the historical production observation data of the reservoir, resulting in: ; In the formula, This represents the actual observed vector of the reservoir. For a moment, , This represents the total reservoir production time. For production well serial number, , Total number of producing wells; This refers to the serial number of the water injection well. , This represents the total number of injection wells. for Time of the first Measured oil production of a production well, in units of ; for Time of the first Measured bottomhole flowing pressure of a production well, in units of ; for Time of the first The measured injection pressure of the injection well, in units of ; This is the transpose of the matrix.

[0047] Step 3.2: Set up a Markov chain noise mechanism and construct a diffusion model to simulate the forward noise diffusion process, thereby smoothly transforming the actual reservoir observation vector into noisy parameters and generating a noisy logarithmic parameter vector, resulting in: ; In the formula, The conditional probability distribution of the forward noise diffusion process is used to characterize the process from... arrive The noise addition process is dimensionless; The time step number is dimensionless. , The total number of diffusion time steps, dimensionless; For the first The logarithmic parameter vector after adding noise is dimensionless. For the first The logarithmic parameter vector after adding noise is dimensionless. It is a multivariate normal distribution, dimensionless, and in this embodiment it is set to a standard normal distribution; For the first The noise schedule coefficient of the step is dimensionless and is a positive real number; It is an identity matrix, dimensionless.

[0048] Step 3.3: Construct an inversion model for inter-well connectivity parameters based on the graph-conditional dual-path diffusion inversion framework, such as... Figure 3 As shown, it includes a conditional inverse denoising network for fusion production data encoders, a multi-condition fusion layer, and a conductivity-volume physics decoupled dual-path denoising network.

[0049] Based on the reservoir historical production observation data, output conditional feature vectors are generated, and then a graph conditional denoising network based on the inter-well connectivity topology is constructed to generate an inter-well connectivity topology map. After multi-condition fusion of the noisy parameter vector, conditional feature vector, and inter-well connectivity topology map, the data is input into the conductivity-volume physical decoupled dual-path denoising network to invert and obtain the posterior samples of inter-well connectivity parameters, including the equivalent conductivity and equivalent connectivity volume of each effective connection between wells.

[0050] Furthermore, the production data encoder is equipped with four one-dimensional convolutional layers and a Transformer self-attention layer, which are used to extract conditional feature vectors from historical reservoir production observation data, capture long-range temporal dependencies and core features in historical reservoir production observation data, input historical reservoir production observation data into the production data encoder, and use the production data encoder to process and output conditional feature vectors.

[0051] The graph-conditional denoising network is set as an inverse denoising network, using GAT as the backbone, so that the GPSNet inter-well connectivity parameter updates of each effective connection between wells can perceive the spatial dependencies of neighboring injection and production wells and adjacent connections, and obtain the inter-well connectivity topology map, resulting in: ; in, ; ; In the formula, This is a dimensionless topological graph showing interconnectedness between wells. For the set of well nodes, dimensionless; This is a dimensionless set of injection well nodes. This is a dimensionless set of production well nodes. This is the set of effective connecting edges between wells, and is dimensionless. This is the serial number of the effective connection between wells, dimensionless.

[0052] The first Logarithmic parameter vector after step-by-step noise addition Based on physical properties, they are divided into logarithmic subvectors of conductivity. and connected volume logarithmic subvector and will the After adding noise, the noisy conductivity component and noisy volume component of the effective connection between each well are used as edge features. Combined with the node features of each injection well node and each production well node, the standardized encoding is obtained as follows: ; In the formula, For the first The initial edge feature vector of the effective connection between the strips is dimensionless; , The first After adding noise, the first step The logarithmic components of the noise-containing conductivity and the logarithmic components of the noise-containing connected volume of the effective connection between the wells are dimensionless. For nodes The initial node feature vectors are dimensionless; The node feature encoding mapping is dimensionless; For nodes The historical production statistics feature vector of the corresponding well; The conditional feature vector output by the production data encoder is dimensionless. ,in, To produce data encoders, To generate network parameters for the data encoder, This represents the actual observed vector of the reservoir. For diffusion time step The positional encoding is dimensionless.

[0053] Inverse denoising is performed using a conductivity-volume physical decoupled dual-path denoising network with node-edge joint message passing. The inverse denoising process is as follows: ; In the formula, This represents the conditional probability distribution for the reverse denoising process; It is the mean function of the inverse denoised distribution and is dimensionless; For the first The variance coefficient of the inverse denoising distribution after step-by-step noise addition is dimensionless.

[0054] By aggregating the information of adjacent edges and adjacent nodes, the edge feature vectors and node feature vectors of each effective connection between wells are updated to obtain the updated edge feature vectors and node feature vectors of each effective connection between wells. This ensures that the update of the effective connection between wells satisfies the spatial continuity of injection and production wells and the correlation of local seepage.

[0055] The update process for the edge feature vectors and node feature vectors of the effective connections between wells is as follows: ; ; In the formula, The layer number is a dimensionless number representing the network layer. , All are node serial numbers; , They are nodes and nodes In the The node representation of a layer in a layered graph network is dimensionless. It is a non-linear activation function, dimensionless; For nodes Adjacent connected sets, dimensionless; , The first Layered graph network layer, first The first layer in a layered graph network The graph attention weights for effective connections between wells are dimensionless. The message transformation matrix is ​​dimensionless. The mapping is updated for edge features and is dimensionless. , The first In a layered graph network, the layer located at the first Effective connection between the two ends of the injection well node between the strip wells Production well nodes The node representation is dimensionless.

[0056] The conductivity-volume physical decoupled dual-path denoising network includes a conductivity network configured in parallel. and volume network The intermediate hidden representation of the conductivity network is obtained. Intermediate hidden representation of volumetric networks , , ,in, Let be the set of network parameters for a conductivity network, which is dimensionless. is the set of network parameters for a volumetric network, which is dimensionless.

[0057] To avoid completely separating the two types of parameters, in the conductivity network and volume network A cross-attention fusion layer is introduced into the intermediate layer to allow conductivity information and volume information to be mutually corrected, which is used to express the synergistic effect of conductivity and reservoir supply capacity in seepage mechanics, resulting in: ; ; In the formula, This is the hidden representation of the conductance network after cross-attention fusion; This represents the hidden representation of the volumetric network after cross-attention fusion; The attention computation operator is dimensionless. , , , , , All are learnable linear mapping matrices, dimensionless.

[0058] For conductivity networks Predicted conductance noise and volumetric networks Predicted volume noise The predicted noise is obtained by splicing. , ,in, This is a vector concatenation operation, dimensionless.

[0059] The predicted noise is obtained after processing using a conductivity-volume physical decoupled dual-path denoising network. , Furthermore, the predicted noise is used as an explicit condition for the inverse denoising distribution. Compared to the conditional diffusion models commonly used in the prior art, which are constructed by directly processing parameter vectors based on U-Net networks or multilayer perceptrons, this embodiment achieves better spatial consistency and seepage physical rationality in the inversion results by explicitly encoding the GPSNet well-to-well connectivity prior.

[0060] Step 3.4: Construct the multi-constraint loss function for the well connectivity parameter inversion model.

[0061] The multi-constraint loss function is set as follows: ; in, ; ; ; ; In the formula, The value of the multi-constraint loss function is dimensionless. The noise prediction loss for the conductivity network is dimensionless. The noise prediction loss for the volumetric network is dimensionless. The weighting coefficients for the interval constraint loss are dimensionless. The interval-constrained soft loss is logarithmic in parameter form and is dimensionless. The weighting coefficient for the physical coupling consistency loss is dimensionless. The loss is a dimensionless loss due to physical coupling consistency. For expectation operations, dimensionless; The true logarithmic parameter vector without noise; This is the number of the diffusion time step, dimensionless; , is the true Gaussian noise component corresponding to the logarithmic subvector of conductivity, which is dimensionless; Transitivity noise predicted by the transitivity network; The time steps of the diffusion process are numbered, and are dimensionless. The conditional feature vector output by the production data encoder is dimensionless. The true Gaussian noise component corresponding to the logarithmic subvector of the connected volume is dimensionless. Volume noise predicted by the volume network; This refers to the 2-norm square operation. The index of the logarithmic parameter vector; This represents the total number of effective connections between wells. It is a function for maximizing the value; For the first A logarithmic parameter vector; For the first The upper bound of the final constraint interval of the logarithmic parameter vector is dimensionless. For the first The lower bound of the final constraint interval for the logarithmic parameter vector is dimensionless. This is a normalized weight matrix for observations, used to balance different magnitudes of production indicators; it is dimensionless. Forward mapping model of GPSNet; For the first The noisy parameters after adding noise and the noise-free logarithmic parameter estimates obtained by restoring the predicted noise are dimensionless. For reservoir history well control vectors; This represents the actual observed vector of the reservoir.

[0062] Compared to traditional diffusion models that uniformly process all parameters, this invention reduces the difficulty of a single network simultaneously learning heterogeneous physical laws through physical decoupling, and maintains the physical synergy between conductivity and volume through cross-attention.

[0063] Step 4: Train the well connectivity parameter inversion model using a standardized training dataset to obtain the trained well connectivity parameter inversion model. This model is used to accurately learn the posterior distribution of well connectivity parameters. The trained well connectivity parameter inversion model is then used to generate logarithmic parameter vector posterior samples and perform uncertainty quantification to construct the logarithmic parameter vector posterior sample set. This includes the following sub-steps: Step 4.1: Train the well connectivity parameter inversion model using a standardized training dataset. During training, implement posterior iterative shrinkage to improve the inversion accuracy of the well connectivity parameter inversion model. After the well connectivity parameter inversion model completes each round of training and inversion, it generates a dataset containing... The posterior sample set of the posterior samples of the inter-well connectivity parameters is used to take the logarithmic parameter vectors in the posterior sample set. The parameter interval is iteratively shrunk by the quantile interval to obtain the trained well connectivity parameter inversion model, and the updated quantile interval is used as the parameter constraint.

[0064] Specifically, the parameter interval iterative shrinkage process is as follows: ; In the formula, For the first The upper bound of the final constraint interval of the logarithmic parameter vector is dimensionless. For the first The lower bound of the final constraint interval for the logarithmic parameter vector is dimensionless. for Quantile operations are dimensionless. For the posterior sample of all well connectivity parameters, the first... The set of values ​​for a logarithmic parameter vector, dimensionless; The quantile threshold is dimensionless. .

[0065] In this embodiment, posterior iterative shrinkage is implemented to improve inversion accuracy. After the first round of training, 200 posterior samples of inter-well connectivity parameters are generated, and the quantile threshold is set. The quantile interval was set to 0.05 and updated. A second round of training was conducted. By iteratively narrowing the interval, the inverted well connectivity parameters were made closer to the actual geology.

[0066] Step 4.2 employs a differential implicit accelerated sampling algorithm, utilizing the well connectivity parameter inversion model to accelerate the generation of posterior samples for well connectivity parameters. It initializes the noisy logarithmic parameter vector based on a standard normal distribution, and simultaneously inputs the current diffusion time step, the noisy parameter vector, the conditional feature vector, and the well connectivity topology map at each inverse diffusion step. The trained well connectivity parameter inversion model is then used to predict noise, and denoising is performed progressively according to the inverse diffusion time step. This is combined with the DDIM accelerated sampling algorithm to reduce the number of sampling steps and improve sampling efficiency, generating... The standard input variables are set and uncertainty quantification analysis is performed to construct a logarithmic parameter vector posterior sample set.

[0067] The logarithmic parameter vector includes the equivalent conductivity and equivalent connectivity volume of the effective interconnectivity between wells. The inversion uncertainty quantification results of the equivalent conductivity and equivalent connectivity volume are as follows: Figure 4 and Figure 5 As shown, based on the posterior samples of multiple sets of logarithmic parameter vectors, the distribution range and discrete characteristics of the connectivity strength of effective connections between wells are statistically analyzed. The results show that the distribution range of equivalent conductivity and equivalent connected volume obtained by inversion can accurately cover the true values, verifying the inversion accuracy and uncertainty quantification capability of the well connectivity parameter inversion model of this invention, and providing a reliable parameter basis for the subsequent construction of representative scenario sets and robust optimization.

[0068] Step 5: Select log-parameter vector posterior samples from the log-parameter vector posterior sample set to construct a representative scenario set to support subsequent optimization decisions.

[0069] In this embodiment, logarithmic parameter vector posterior samples are selected from the logarithmic parameter vector posterior sample set. The degree of fit between the logarithmic parameter vector posterior samples and the historical reservoir observation data is quantified using a mismatch function. A preset fitting threshold of 0.05 is used. Logarithmic parameter vector posterior samples with mismatch function values ​​not exceeding the fitting threshold are retained as valid logarithmic parameter vector posterior samples. Invalid logarithmic parameter vector posterior samples are then removed from the logarithmic parameter vector posterior sample set, resulting in a set of 100 valid posterior samples. .

[0070] The mismatch function is set as follows: ; In the formula, For the first The reservoir history fitting mismatch function value of the group of effective posterior sample sets; The weighting coefficients for the yield fit are dimensionless. For the first The simulated output vector obtained from the effective posterior sample set through the GPSNet forward mapping model is dimensionless; This is the field-measured vector of reservoir production, dimensionless. For 2-norm square operations; for; For the first The simulated pressure vector obtained from the effective posterior sample set through the GPSNet forward mapping model is dimensionless; The pressure is a dimensionless vector measured in the field. This represents the total number of valid posterior samples.

[0071] Then construct a mean parameter model, and use the mean parameter model to analyze each effective posterior sample set. Perform mean calculation to determine the logarithmic parameter mean vector. The calculation formula is: ; In the formula, This is a vector of logarithmic parameter means, including the mean of the equivalent conductivity of the effective connections between each well. and equivalent connected volume ; For the effective logarithmic parameter vector posterior samples in the effective posterior sample set.

[0072] Then, by inverse logarithmic transformation, the mean equivalent conductivity and mean equivalent connectivity volume of the effective connections between each well are obtained, resulting in: ; In the formula, For the first Mean equivalent conductivity of effective connections between wells; For the first The baseline value of the equivalent conductivity of the effective connection between wells; For the first Mean equivalent conductivity of effective connections between wells; For the first Mean equivalent connectivity volume of effective connections between wells; For the first The baseline value of the equivalent connectivity volume for effective connections between wells; For the first The equivalent connected volume of the effective connection between the strip wells; It is the Euler number.

[0073] Finally, the K-means hierarchical clustering algorithm was used to analyze each effective posterior sample set. Cluster analysis was performed by using the effective posterior sample set The effective logarithmic parameter vector in the posterior sample is divided into _ ... For each category, select the posterior sample of the effective log parameter vector closest to the cluster center to construct a representative scene set. , , Representative scene set The Middle The posterior sample of the effective log-parameter vector that is closest to the cluster center in each category.

[0074] This embodiment uses production wells P1 and P2 as examples to demonstrate the fitting effect of historical reservoir production data corresponding to representative scenario sets, such as... Figure 6 and Figure 7 As shown. The water-drive reservoirs described in production wells P1 and P2 are developed using an injection-production balance system. The average daily production of the production wells in the water-drive reservoir is 13... The average daily water injection volume of the injection well is 18 , Figure 6 and Figure 7 In the diagram, the light green area represents the fitting interval for 10 representative scenarios, while the solid line represents the actual measured curve.

[0075] analyze Figure 6 and Figure 7 As can be seen, the measured curves of daily oil production and water cut of the two wells are completely enveloped by the fitted interval. The daily oil production smoothly transitioned from high production capacity in the early stage to stable oil production level in the later stage, which fully verified the high consistency between the representative scenario and the historical production data of the reservoir, as well as the physical rationality of the representative scenario under the injection-production balance system, and provided reliable support for solving the subsequent robust optimization injection-production scheme.

[0076] Step 6: Obtain the reservoir development scheme to be optimized, construct a dual-mode reservoir optimization model including a deterministic optimization model and a robust optimization model, solve the dual-mode reservoir optimization model in representative scenarios of a representative scenario set based on intelligent optimization algorithms, obtain the uncertainty propagation quantification index of the dual-mode reservoir optimization model, and determine whether it exceeds the preset trigger threshold. If it exceeds the preset trigger threshold, trigger the prior constraint dynamic feedback adjustment mechanism; otherwise, output the optimized reservoir development scheme, including the following steps: Step 6.1: Obtain the reservoir development scheme to be optimized and determine the decision variables of the reservoir development scheme to be optimized. And in conjunction with the engineering physical feasible region constraints that determine the decision variables. A net present value objective function is constructed using the reservoir net present value as the development objective. .

[0077] In this embodiment, the engineering physical feasible domain constraints include single-well limit injection capacity constraints, total water injection capacity constraints of the gathering and transportation station, fluid discharge capacity constraints of the production well lifting equipment, and stability constraints of injection volume distribution in adjacent time periods. Furthermore, the upper limit for single-well injection is set to 40. The total water injection limit is set at 200. The production well's maximum fluid production is set at 200. The phased injection fluctuation does not exceed 20%, and the injection volume in the first optimized period is seamlessly connected with the historical end-of-period water injection status, thus ensuring the field feasibility of the reservoir development plan optimization.

[0078] Specifically, using net present value (NPV) as the core development objective function, which balances production revenue and development costs, and discounting future revenues using a discount rate over a discrete forecast period, a quantitative assessment of the economic benefits of reservoir development schemes is achieved, thus constructing the NPV objective function. .

[0079] The net present value objective function Set to: ; In the formula, Representative scene set The Middle The posterior sample of the effective log-parameter vector closest to the cluster center in each category; This refers to the time step number; This represents the total number of time steps. The profit coefficient per unit of oil production, in units of ; for Time of the first Simulated oil production for representative scenarios, in units of ; This is the cost coefficient per unit of water injection, in units of... ; for Time of the first Simulated water injection volume for a representative scenario, in units of ; The cost coefficient per unit of production operation, in units of ; for Time of the first Simulated production operation volume for a representative scenario, in units of ; The annual discount rate is dimensionless. For time step The corresponding absolute time, in days; This is the time scale conversion factor.

[0080] In this embodiment, the profit coefficient per unit oil production Set to 2200 Cost coefficient per unit volume of water injected Set to 35 Cost coefficient per unit of production operation Set to 18 Annual discount rate Set to 0.08, time scale conversion factor Set it to 365.

[0081] Step 6.2: Construct a dual-mode optimization model for the reservoir, including a deterministic optimization model and a robust optimization model, wherein the deterministic optimization model is set as follows: , The maximum value function is defined as follows: The robust optimization model is set as follows: ,in, , , Decision variables Expected returns in representative scenarios, expressed in yuan; The serial number represents the scene. The total number of representative scenarios; Decision variables The variance of returns in a representative scenario is used to characterize the degree of return volatility, and the unit is 1 / 2 oz. ; The risk aversion coefficient is dimensionless.

[0082] Step 6.3: In all representative scenarios of the representative scenario set, an intelligent optimization algorithm is used to optimize the decision variables within the engineering-physical feasible region constraints. Solve deterministic and robust optimization models in the middle.

[0083] In this embodiment, the intelligent optimization algorithm used is specifically set as the particle swarm optimization algorithm, which maps the decision variables to particle position vectors in a high-dimensional search space. That is, the net present value objective function determined by the particle position in the particle swarm and the particle fitness are the decision variables. Set the total number of particles in the particle swarm to 50 and the total number of iterations to [number missing]. For 200 iterations, within the engineering physical feasible region constraint. The positions and velocities of each particle in the particle swarm are initialized. The particle swarm optimization algorithm is used to solve the deterministic optimization model and the robust optimization model respectively, and the optimal decision variables obtained by using the deterministic optimization model and the robust optimization model are obtained.

[0084] Specifically, in the particle swarm optimization algorithm, the particle velocity and position update process is as follows: ; ; In the formula, The number of iterations for particle swarm optimization, dimensionless; For the first The particle velocity vector of the next iteration is dimensionless. For the first The particle velocity vector of the next iteration is dimensionless. This is the inertia weighting coefficient. ; As an individual cognitive learning factor, ; , All Random numbers within the interval, dimensionless; As a social cognitive learning factor, ; For the first The optimal position of each individual particle; The optimal position for the entire population; For the first The particle position in the next iteration; For the first The particle position in the next iteration.

[0085] In calculating the particle fitness, the decision variables corresponding to each particle are substituted into all representative scenarios of the representative scenario set to conduct reservoir numerical simulation. The net present value of each representative scenario is calculated, and the minimum value is taken as the particle fitness value. During the iterative optimization process, a boundary absorption strategy is introduced, and the upper limit of single-well injection volume is set to 30 based on the actual production constraints of the oilfield. The lower limit for single-well injection volume is set at 10. For particles whose positions exceed the updated physical boundary, a forced projection correction is performed to bring them back to the nearest feasible boundary, ensuring that all candidate solutions meet the injection and extraction process requirements. Meanwhile, in this embodiment, the particle swarm optimization algorithm's termination condition is set as a dual iterative convergence criterion; when the total number of iterations... When the number of particles is greater than 200 or the change in the global optimal fitness of the particle swarm over 20 consecutive generations is less than 10⁻³, the particle swarm optimization ends, the optimized reservoir development scheme is output, the injection volume decision in the optimized reservoir development scheme is determined, and the optimal robust injection and production scheme adapted to multiple operating conditions is obtained.

[0086] Figure 8 This is a schematic diagram of the convergence curve during the optimization process of the particle swarm optimization algorithm in this embodiment. Figure 8 As can be seen, the particle swarm optimization algorithm in this embodiment converges after 140 iterations, with the expected net present value stabilizing at 5.6 million yuan. The worst-case net present value also converges to 5 million yuan, and the revenue variance gradually converges. This verifies the efficiency of the particle swarm optimization algorithm and the stability of the robust optimization model used in this embodiment, providing support for solving the optimal reservoir development scheme.

[0087] After particle swarm optimization, uncertainty propagation quantification metrics are obtained for both the deterministic and robust optimization models, transforming the abstract uncertainty propagation into quantifiable numerical indicators. These uncertainty propagation quantification metrics include control drift metrics. and return dispersion index , , ,in, For the first The optimal decision variable matrix obtained by individual optimization in a representative scenario is dimensionless. This is the mean matrix of the optimal decision matrices for all representative scenarios. ; This is the 2-norm square operation.

[0088] Step 6.4: Based on the uncertainty propagation quantification index of the deterministic optimization model and the robust optimization model, using the uncertainty propagation quantification index of the deterministic optimization model as the standard value, calculate the difference in control drift index and the difference in return dispersion index between the robust optimization model and the deterministic optimization model, such as... Figure 9 and Figure 10As shown, the difference in control drift index and the difference in revenue dispersion index are compared with preset trigger thresholds. If they exceed the preset trigger thresholds, the prior constraint dynamic feedback adjustment mechanism is triggered. The uncertainty propagation quantification index of the robust optimization model is used as the feedback signal. The prior constraint interval, sensitivity interval scaling factor, and prior envelope rate are adjusted according to the uncertainty propagation quantification index of the robust optimization model. Steps 1 to 6 are re-executed until the prior constraint dynamic feedback adjustment mechanism is no longer triggered, and the optimized reservoir development scheme is output. If the preset trigger thresholds are not exceeded, the prior constraint dynamic feedback adjustment mechanism is not triggered, and the optimized reservoir development scheme is directly output.

[0089] analyze Figure 9 and Figure 10 It can be seen that the control drift of the robust optimization model is reduced by 64% and the revenue dispersion is reduced by 72% compared with the deterministic optimization model. This verifies the significant advantages of the reservoir development scheme obtained by the robust optimization model in suppressing the propagation of parameter uncertainty and improving decision stability, and provides a quantitative basis for judging the necessity of robust optimization.

[0090] Specifically, in this embodiment, if the difference in control drift index between the robust optimization model and the deterministic optimization model exceeds a preset trigger threshold, the prior constraint dynamic feedback adjustment mechanism is triggered, and the interval scaling factor of the effective connection logarithmic parameter vector between the main wells is adjusted. Inter-well effective connection logarithmic parameter vector interval scaling factor Interval scaling factor of the logarithmic parameter vector of effective connection between weak wells The parameters are increased by 0.1, 0.2, and 0.3 respectively to broaden the prior constraint interval. If the difference in the revenue dispersion index between the robust optimization model and the deterministic optimization model exceeds the preset trigger threshold, the prior constraint dynamic feedback adjustment mechanism is triggered. The contraction factors of the high-sensitivity logarithmic parameter vector and the medium-sensitivity logarithmic parameter vector are lowered by 0.1 to 0.2 to strengthen the contraction of the high-sensitivity parameter interval. The prior envelope rate is adjusted to 0.95 to improve the coverage accuracy of the parameter interval on the measured data. Based on the corrected adaptive prior constraints, steps 1 to 6 are re-executed until the prior constraint dynamic feedback adjustment mechanism is no longer triggered. The optimized reservoir development scheme is output, thus realizing a closed-loop decision-making process that goes through adaptive prior constraints, conditional diffusion inversion, sampled robust optimization, and uncertainty feedback.

[0091] In summary, the method of this invention is an integrated reservoir development scheme optimization method that takes into account physical rationality, efficient uncertainty quantification, and low-cost robust decision-making. It realizes accurate inversion of inter-well connectivity parameters and robust optimization of reservoir development schemes, providing technical support for the refined and intelligent development of water-drive reservoirs.

[0092] 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 decision-making method for well connectivity parameter inversion based on a diffusion model, characterized in that, Includes the following steps: Step 1: Based on the actual well network layout of the target injection-production block, select the parameters to be inverted, construct a global parameter vector and perform logarithmic reparameterization to obtain the standard input variables; Step 2: Perform parameter baseline value calculation, hierarchical initial interval setting, dynamic envelope test and sensitivity hierarchical shrinkage on the standard input variable in sequence, adaptively optimize the value range of the standard input variable to obtain prior samples, and construct a standardized training dataset; Step 3: Based on the graph-conditional dual-path diffusion inversion framework that integrates production data, well connectivity topology, and physical properties, construct the well connectivity parameter inversion model and determine the multi-constraint loss function of the well connectivity parameter inversion model. Step 4: Train the well connectivity parameter inversion model using the standardized training dataset to obtain the trained well connectivity parameter inversion model, which is used to accurately learn the posterior distribution of well connectivity parameters. Use the trained well connectivity parameter inversion model to generate log parameter vector posterior samples and perform uncertainty quantification to construct a log parameter vector posterior sample set. Step 5: Select log-parameter vector posterior samples from the log-parameter vector posterior sample set to construct a representative scenario set to support subsequent optimization decisions; Step 6: Obtain the reservoir development scheme to be optimized, construct a dual-mode reservoir optimization model including a deterministic optimization model and a robust optimization model, solve the dual-mode reservoir optimization model in each representative scenario of the representative scenario set based on the intelligent optimization algorithm, obtain the uncertainty propagation quantification index of the dual-mode reservoir optimization model, and determine whether it exceeds the preset trigger threshold. If it exceeds the preset trigger threshold, trigger the prior constraint dynamic feedback adjustment mechanism; otherwise, output the optimized reservoir development scheme.

2. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1.1: Based on the actual well network layout of the target injection-production block, obtain the total number of effective connections between wells within the target injection-production block. The equivalent conductivity and equivalent connectivity volume of each effective connection between wells are selected as parameters to be inverted. A global parameter vector is constructed to characterize the inter-well connectivity features of the target injection-production block, resulting in: ； In the formula, This is the global parameter vector; For the real number field; For the first Equivalent conductivity of effective connections between wells; For the first The equivalent connected volume of the effective connection between the strip wells; It is the transpose matrix; Step 1.2 introduces logarithmic reparameterization to eliminate training instability caused by the order-of-magnitude difference in the parameters to be inverted, and pre-sets the equivalent transmissivity benchmark value of the effective interconnection between wells based on the reservoir geology prior of the target injection-production block. and equivalent connected volume benchmark value By performing a logarithmic transformation on the global parameter vector and reconstructing the standard input variables, we obtain: ； In the formula, For standard input variables; For the first Log-reparameterized value of the equivalent conductivity of the effective connection between wells; For the first Logarithmic reparameterized value of the equivalent connected volume of the effective connection between wells.

3. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1, for the effective connection between wells within the target injection-production block, determine the equivalent conductivity benchmark value based on the principles of seepage mechanics: ； In the formula, For the first The baseline value of the equivalent conductivity of the effective connection between wells; For the first Unit conversion and network equivalence correction coefficients for effective connections between wells; For the first Equivalent permeability index of effective connection between wells; For the first The equivalent flow cross-sectional area of ​​the effective connection between the shafts; For the first Equivalent fluid viscosity of the effective connection between wells; For the first The equivalent length of the effective connection between the strip wells; Based on reservoir volume characteristics, the equivalent connectivity volume benchmark value for effective connections between wells is determined as follows: ； In the formula, For the first The baseline value of the equivalent connectivity volume for effective connections between wells; For the first Volumetric equivalent correction factor for effective connection between shafts; For the first Equivalent porosity of effective connection between wells; For the first The equivalent thickness of the effective connection between the strip wells; Step 2.2: Based on the pre-defined prior constraint intervals according to the strength classification of inter-well connectivity, the effective connections between each well are divided into a primary connectivity candidate set, a normal connectivity candidate set, and a weak connectivity candidate set. The threshold values ​​for each type of connectivity candidate set in the prior constraint interval are set as follows: ； In the formula, For the first The lower bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells; For the first The upper bound of the prior constraint interval for the logarithmic parameter vector of the equivalent conductivity of the effective connection between wells; The serial number of the effective connection between wells; Principal connected candidate set; The interval scaling factor of the logarithmic parameter vector of the effective connection between main wells; This is a candidate set for ordinary connectivity. For the interval scaling factor of the logarithmic parameter vector of the effective connection between ordinary wells; A candidate set of weakly connected networks; The interval scaling factor for the logarithmic parameter vector of the effective connection between weak wells; ; Step 2.3: Optimize the prior constraint interval based on the dynamic envelope test, set the envelope indicator function and the prior envelope rate, and use Latin hypercube sampling to draw from the prior constraint interval. A set of prior samples are input into the GPSNet forward model to obtain simulated production vectors, which are then used as historical production observation data for the reservoir. Step 2.4: Perform parametric sensitivity analysis on each logarithmic parameter vector in the standard input variables to complete the graded shrinkage. Calculate the roughness sensitivity of each logarithmic parameter vector. Based on the roughness sensitivity of each logarithmic parameter vector in the standard input variables, classify each logarithmic parameter vector into high sensitivity, medium sensitivity, and low sensitivity. Determine the differential shrinkage factor based on the sensitivity of each logarithmic parameter vector. Determine the shrunken prior interval, obtain the filtered prior parameter samples, and construct a standardized training dataset. The formula for calculating roughness sensitivity is: ； In the formula, For standard input variables The Middle The coarseness sensitivity of a logarithmic parameter vector; This refers to the prior sample group number; The first in the standard input variables Group of prior samples; The perturbation scale; For the first Unit vectors; For reservoir history well control vectors; For 2-norm operations; This is the GPSNet forward mapping model.

4. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Using historical production observation data of the reservoir as a condition and learning the posterior distribution of the logarithmic parameter vector as the inversion target, obtain the actual observation vector of the reservoir, resulting in: ； In the formula, This represents the actual observed vector of the reservoir. For a moment, , This represents the total reservoir production time. For production well serial number, , Total number of producing wells; This refers to the serial number of the water injection well. , This represents the total number of injection wells. for Time of the first Measured oil production of a production well; for Time of the first Measured bottomhole flowing pressure of a production well; for Time of the first Measured injection pressure of the injection well; It is the transpose matrix; Step 3.2: Set up a Markov chain noise-adding mechanism, construct a diffusion model to simulate the forward noise-adding diffusion process, and generate a noisy logarithmic parameter vector, resulting in: ； In the formula, This represents the conditional probability distribution of the forward noise diffusion process; This is the numbering of the diffusion time step. , This represents the total number of diffusion time steps. For the first The logarithmic parameter vector after adding noise; For the first The logarithmic parameter vector after adding noise; It follows a multivariate normal distribution; For the first The noise schedule factor of the step; It is the identity matrix; Step 3.3: Based on the graph conditional dual-path diffusion inversion framework, construct the well connectivity parameter inversion model, including a conditional reverse denoising network that fuses production data encoders, a multi-condition fusion layer, and a conductivity-volume physics decoupled dual-path denoising network; Based on the reservoir historical production observation data, output conditional feature vectors are generated, and then a graph conditional denoising network based on the inter-well connectivity topology is constructed to generate an inter-well connectivity topology map. After multi-condition fusion of the noisy parameter vector, conditional feature vector, and inter-well connectivity topology map, the data is input into the conductivity-volume physical decoupled dual-path denoising network to invert and obtain the posterior samples of inter-well connectivity parameters, including the equivalent conductivity and equivalent connectivity volume of each effective connection between wells. Step 3.4: Construct the multi-constraint loss function for the well connectivity parameter inversion model.

5. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 4, characterized in that, The production data encoder is equipped with four one-dimensional convolutional layers and a Transformer self-attention layer, which are used to extract conditional feature vectors from historical reservoir production observation data, capture long-range temporal dependencies and core features in historical reservoir production observation data, input historical reservoir production observation data into the production data encoder, and use the production data encoder to process and output conditional feature vectors. The graph conditional denoising network is set as an inverse denoising network, using GAT as the backbone, to obtain the well connectivity topology graph, resulting in: ； in, ； ； In the formula, This is a topology diagram showing the connectivity between wells; For the set of well nodes; For the set of water injection well nodes; For the set of production well nodes; This is the set of effective connecting edges between wells; The serial number of the effective connection between wells; The first Logarithmic parameter vector after step-by-step noise addition Based on physical properties, they are divided into logarithmic subvectors of conductivity. and connected volume logarithmic subvector and will the After adding noise, the noisy conductivity component and noisy volume component of the effective connection between each well are used as edge features. Combined with the node features of each injection well node and each production well node, the standardized encoding is obtained as follows: ； In the formula, For the first The initial edge feature vector of the effective connection between the strips; , The first After adding noise, the first step Logarithmic component of noisy conductivity and logarithmic component of noisy connectivity volume in the effective connection between wells; For nodes The initial node feature vector; Encode and map node features; For nodes The historical production statistics feature vector of the corresponding well; To produce the conditional feature vector output by the data encoder, ,in, To produce data encoders, To generate network parameters for the data encoder, This represents the actual observed vector of the reservoir. For diffusion time step Location encoding.

6. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 5, characterized in that, Inverse denoising is performed using a conductivity-volume physical decoupled dual-path denoising network with node-edge joint message passing. The inverse denoising process is as follows: ； In the formula, This represents the conditional probability distribution for the reverse denoising process; It is the mean function of the inverse denoising distribution; For the first The variance coefficient of the inverse denoising distribution after step-by-step noise addition; By aggregating adjacent edges and neighboring node information, the edge feature vectors of effective connections between each well and the feature vectors of each node are updated to obtain the updated edge feature vectors of effective connections between each well and the feature vectors of each node. The feature vector update process is as follows: ； ； In the formula, The layer numbering is used for the graph network. , All are node serial numbers; , They are nodes and nodes In the Node representation of a layer in a layered graph network; It is a non-linear activation function; For nodes Adjacent connection sets; , The first Layered graph network layer, first The first layer in a layered graph network Graph attention weights for effective connections between wells; The message transformation matrix; Update the mapping for edge features; , The first In a layered graph network, the layer located at the 1st Effective connection between the two ends of the injection well node between the strip wells Production well nodes The node representation; The conductivity-volume physical decoupled dual-path noise reduction network includes a conductivity network configured in parallel. and volume network The intermediate hidden representation of the conductivity network is obtained. Intermediate hidden representation of volumetric networks , , ,in, For the set of network parameters of a conductivity network, This is the set of network parameters for the volumetric network. In conductivity network and volume network A cross-attention fusion layer is introduced into the intermediate layer to allow conductivity information and volume information to be mutually corrected, which is used to express the synergistic effect of conductivity and reservoir supply capacity in seepage mechanics, resulting in: ； ； In the formula, This is the hidden representation of the conductance network after cross-attention fusion; This represents the hidden representation of the volumetric network after cross-attention fusion; Operators are computed for attention; , , , , , All are learnable linear mapping matrices; For conductivity networks Predicted conductance noise and volumetric networks Predicted volume noise The predicted noise is obtained by splicing. , ,in, This is a vector concatenation operation; The predicted noise is obtained after processing using a conductivity-volume physical decoupled dual-path denoising network. , Furthermore, the predicted noise is used as an explicit condition for the inverse denoising distribution.

7. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 4, characterized in that, The multi-constraint loss function is set as follows: ； in, ； ； ； ； In the formula, For the multi-constraint loss function value; Predict the noise loss for the conductivity network; Predict the noise loss for volumetric networks; These are the weighting coefficients for the interval constraint loss; For interval-constrained soft loss with logarithmic parameters; The weighting coefficients for the physical coupling consistency loss; This represents a loss of consistency due to physical coupling. For expectation calculation; The true logarithmic parameter vector without noise; This is the number of the diffusion time step; This represents the true Gaussian noise component corresponding to the logarithmic subvector of conductivity. Transitivity noise predicted by the transitivity network; Number the time steps of the diffusion process; The conditional feature vector output by the production data encoder; The true Gaussian noise component corresponding to the logarithmic subvector of the connected volume; Volume noise predicted by the volume network; This refers to the 2-norm square operation; The index of the logarithmic parameter vector; This represents the total number of effective connections between wells; It is a function for maximizing the value; For the first A logarithmic parameter vector; For the first The upper bound of the final constraint interval for the logarithmic parameter vector; For the first The lower bound of the final constraint interval for the logarithmic parameter vector; The normalized weight matrix for the observations; Forward mapping model of GPSNet; For the first The noisy parameters after adding noise and the noiseless logarithmic parameter estimates obtained by restoring the predicted noise.

8. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Train the well connectivity parameter inversion model using a standardized training dataset. During training, implement posterior iterative shrinkage to improve the inversion accuracy of the well connectivity parameter inversion model. After the well connectivity parameter inversion model completes each round of training and inversion, it generates a dataset containing... The posterior sample set of the posterior samples of the inter-well connectivity parameters is used to take the logarithmic parameter vectors in the posterior sample set. The parameter interval is iteratively shrunken by the quantile interval to obtain the trained well connectivity parameter inversion model, and the updated quantile interval is used as the parameter constraint. The parameter interval iterative shrinkage process is as follows: ； In the formula, For the first The upper bound of the final constraint interval for the logarithmic parameter vector; For the first The lower bound of the final constraint interval for the logarithmic parameter vector; for Quantile operations; For the posterior sample of all well connectivity parameters, the first... The set of possible values ​​for a logarithmic parameter vector; This is the quantile threshold; Step 4.2 employs a differential implicit accelerated sampling algorithm, utilizing the well connectivity parameter inversion model to accelerate the generation of posterior samples for well connectivity parameters. It initializes the noisy logarithmic parameter vector based on a standard normal distribution, and simultaneously inputs the current diffusion time step, the noisy parameter vector, the conditional feature vector, and the well connectivity topology map at each inverse diffusion step. The trained well connectivity parameter inversion model is then used to predict noise, and denoising is performed progressively according to the inverse diffusion time step. This is combined with the DDIM accelerated sampling algorithm to reduce the number of sampling steps and improve sampling efficiency, generating... The standard input variables are set and uncertainty quantification analysis is performed to construct a logarithmic parameter vector posterior sample set.

9. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, In step 5, logarithmic parameter vector posterior samples are selected from the logarithmic parameter vector posterior sample set. A mismatch function is used to quantify the fit between the logarithmic parameter vector posterior samples and historical reservoir observation data. Invalid logarithmic parameter vector posterior samples are then removed from the logarithmic parameter vector posterior sample set. Group effective posterior sample set ; The mismatch function is set as follows: ； In the formula, For the first The reservoir history fitting mismatch function value of the group of effective posterior sample sets; These are the weighting coefficients for the yield fit; For the first The simulated output vector obtained from the effective posterior sample set of the group through the GPSNet forward mapping model; This represents the field-measured vector of reservoir production. This refers to the 2-norm square operation; For the first The simulated pressure vector is obtained from the effective posterior sample set through the GPSNet forward mapping model; This is the field-measured vector of pressure; Then construct a mean parameter model, and use the mean parameter model to analyze each effective posterior sample set. Perform mean calculation to determine the logarithmic parameter mean vector. The calculation formula is: ； In the formula, This is a vector of logarithmic parameter means, including the mean of the equivalent conductivity of the effective connections between each well. and equivalent connected volume ; For the effective logarithmic parameter vector posterior samples in the effective posterior sample set; Then, by inverse logarithmic transformation, the mean equivalent conductivity and mean equivalent connectivity volume of the effective connections between each well are obtained, resulting in: ； In the formula, For the first Mean equivalent conductivity of effective connections between wells; For the first The baseline value of the equivalent conductivity of the effective connection between wells; For the first Mean equivalent conductivity of effective connections between wells; For the first Mean equivalent connectivity volume of effective connections between wells; For the first The baseline value of the equivalent connectivity volume for effective connections between wells; For the first The equivalent connected volume of the effective connection between the strip wells; It is the Euler number; Finally, the K-means hierarchical clustering algorithm was used to analyze each effective posterior sample set. Cluster analysis was performed by using the effective posterior sample set The effective logarithmic parameter vector in the posterior sample is divided into _ ... For each category, select the posterior sample of the effective log parameter vector closest to the cluster center to construct a representative scene set. , , Representative scene set The Middle The posterior sample of the effective log-parameter vector that is closest to the cluster center in each category.

10. The well connectivity parameter inversion and robust decision-making method based on the diffusion model according to claim 1, characterized in that, Step 6 includes the following steps: Step 6.1: Obtain the reservoir development scheme to be optimized and determine the decision variables of the reservoir development scheme to be optimized. And in conjunction with the engineering physical feasible region constraints that determine the decision variables. A net present value objective function is constructed using the net present value of the reservoir as the development objective. ; The net present value objective function Set to: ； In the formula, Representative scene set The Middle The posterior sample of the effective log-parameter vector closest to the cluster center in each category; This refers to the time step number; This represents the total number of time steps. The profit coefficient per unit of oil production; for Time of the first Simulated oil production in representative scenarios; Cost coefficient per unit volume of water injected; for Time of the first Simulated water injection volume for a representative scenario; Cost coefficient per unit of production operation; for Time of the first Simulated production operation volume for a representative scenario; The annual discount rate; For time steps The corresponding absolute time; For time scale conversion factors; Step 6.2: Construct a dual-mode optimization model for the reservoir, including a deterministic optimization model and a robust optimization model, wherein the deterministic optimization model is set as follows: , It is a function with maximum value. The mean vector of logarithmic parameters; the robust optimization model is set as follows: ,in, , , Decision variables Expected returns in representative scenarios; The serial number represents the scene. The total number of representative scenarios; Decision variables Variance of returns in representative scenarios; Risk aversion coefficient; Step 6.3: In all representative scenarios of the representative scenario set, an intelligent optimization algorithm is used to optimize the decision variables within the engineering-physical feasible region constraints. Solve deterministic and robust optimization models in the model, and obtain the uncertainty propagation quantification index, including the control drift index, for both models. and return dispersion index , , ,in, For the first The optimal decision variable matrix obtained by individual optimization in a representative scenario; This is the mean matrix of the optimal decision matrices for all representative scenarios. ; This refers to the 2-norm square operation; Step 6.4: Based on the uncertainty propagation quantification index of the deterministic optimization model and the robust optimization model, using the uncertainty propagation quantification index of the deterministic optimization model as the standard value, calculate the difference in control drift index and the difference in revenue dispersion index between the robust optimization model and the deterministic optimization model, and compare them with the preset trigger threshold. If the preset trigger threshold is exceeded, the prior constraint dynamic feedback adjustment mechanism is triggered, using the uncertainty propagation quantification index of the robust optimization model as the feedback signal. The prior constraint interval, sensitivity interval scaling factor, and prior envelope rate are adjusted according to the uncertainty propagation quantification index of the robust optimization model. Steps 1 to 6 are re-executed until the prior constraint dynamic feedback adjustment mechanism is no longer triggered, and the optimized reservoir development scheme is output. If the preset trigger threshold is not exceeded, the prior constraint dynamic feedback adjustment mechanism is not triggered, and the optimized reservoir development scheme is directly output.

Images