A reservoir parameter inversion modeling method integrating surrogate model and online learning
Through the integration of agent model and online learning method, combined with Transformer model and numerical simulation technology, the problem of low inversion efficiency of reservoir parameters is solved, more efficient calculations and more accurate results are achieved, the number of numerical simulations is reduced, and the number of numerical simulations is improved. Applicability of reservoir parameter inversion.
Patent Information
- Application Number
- CN202411497944.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2024-12-31
- Estimated Expiration
- 2044-10-25
Smart Images
Figure CN119005027B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of oil and gas reservoir development, and in particular relates to an oil reservoir parameter inversion modeling method integrating an agent model and online learning. Background Art
[0002] Oil reservoirs are deeply buried underground, with strong uncertainty in geological parameters and complex spatial and temporal scale characteristics. It has always been an important and challenging task to invert the underground parameters of oil reservoirs based on observation data at the oil field site. On the one hand, the actual oil reservoir covers a large area and has strong parameter heterogeneity, while the actual observation data is relatively small in both vertical and planar dimensions, resulting in low accuracy and large uncertainty in the inversion parameters. On the other hand, reservoir parameter inversion often involves hundreds or thousands of numerical simulation iterations, and the actual reservoir model has many grids, which can take days to weeks. Therefore, in actual operations, some accuracy is often sacrificed to increase the inversion speed, which also reduces the reliability of the inversion results. Therefore, there is an urgent need for more efficient reservoir parameter inversion modeling methods to reduce the uncertainty of underground parameters and improve the understanding of underground reservoir geological attribute parameters.
[0003] The current reservoir parameter inversion modeling methods can be mainly divided into two categories: inversion methods based on numerical simulation and inversion methods based on proxy models. The inversion method based on numerical simulation often uses optimization algorithms to iteratively adjust reservoir model parameters, conduct numerical simulation evaluation, calculate the observed data fitting objective function, and repeat the iteration until a satisfactory observed data fitting effect is achieved. The reservoir model at this time is more consistent with the actual reservoir conditions, and the parameters here are the parameters obtained by inversion. However, this method is time-consuming and often requires thousands of numerical simulations. A single numerical simulation of an actual reservoir model can take several hours. In practical applications, this method is inefficient. The inversion method based on the proxy model uses the data generated by the numerical simulation to train a proxy model with fast calculation speed and accuracy close to that of the numerical simulation. It replaces the numerical simulation for evaluation during the iterative process of the optimization algorithm to accelerate the reservoir parameter inversion process. First, it is necessary to combine the numerical simulator to generate a large number of sample libraries of reservoir parameters and corresponding simulated production data, and then build a machine learning proxy model, such as a reduced-order model, radial basis network model, Gaussian process, polynomial regression model, convolutional neural network model, recurrent neural network model, Transformer model, etc. Use sample data to train the proxy model. When the accuracy of the proxy model meets the requirements, it can replace the numerical simulator in the parameter inversion process. Compared with the numerical simulation method, this method has a certain improvement in computational efficiency, but for some complex reservoir scenarios, the nonlinear relationship between reservoir parameters and production data is extremely strong, resulting in a large number of samples required for the training of the proxy model, and the computational cost is further increased, which significantly affects the effect of the proxy model method.
[0004] Therefore, there is an urgent need for a more efficient reservoir parameter inversion modeling method that integrates machine learning proxy models and numerical simulation technology. Summary of the invention
[0005] In order to solve the above problems, the present invention proposes a reservoir parameter inversion modeling method that integrates a proxy model and online learning, which can fully combine the proxy model with numerical simulation technology to more efficiently perform reservoir parameter inversion, reduce parameter uncertainty, reduce the number of reservoir numerical simulations, and improve calculation speed.
[0006] The technical solution of the present invention is as follows:
[0007] A reservoir parameter inversion modeling method integrating a proxy model and online learning comprises the following steps:
[0008] Step 1: Combine the reservoir numerical simulator with prior parameters to obtain reservoir model parameters and simulated production data, and construct a reservoir numerical simulation data set;
[0009] Step 2: construct a reservoir numerical simulation proxy model based on the Transformer model, and train the proxy model in combination with the reservoir numerical simulation data set;
[0010] Step 3: Use the optimization algorithm to iteratively adjust the reservoir model parameter set, and then use the proxy model to evaluate the data fitting effect of the updated model;
[0011] Step 4: In the iteration of the optimization algorithm, an online learning mechanism is constructed to determine the uncertainty of the proxy model prediction based on the random discarding neuron structure and select the reservoir model parameters with large uncertainty;
[0012] Step 5: Perform numerical simulation on the selected reservoir model parameters, obtain new simulated production data, construct a new reservoir numerical simulation data set, and retrain the proxy model using all current data sets;
[0013] Step 6: According to the final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3, the distribution of actual reservoir parameters is determined to complete the reservoir parameter inversion process.
[0014] Furthermore, in step 1, according to the prior parameter sample of the reservoir model, a reservoir numerical simulator is used to perform numerical simulation to obtain reservoir model parameters and corresponding simulated production data; wherein the reservoir model parameters include grid permeability, porosity, and relative permeability, and the simulated production data include oil production, water production, and pressure data.
[0015] Furthermore, in step 2, the reservoir numerical simulation agent model includes a feature extraction module, a feature conversion module, and a time series prediction module; wherein the feature extraction module includes a convolution layer, a downsampling layer, and an activation function; the feature conversion module includes a global pooling layer, a random drop layer, and a replication layer; the time series prediction module includes multiple Transformer layers, each Transformer layer includes a multi-head attention layer, a layer normalization, a linear layer, and also includes an addition operation of cross-layer parameters;
[0016] The specific working process of the reservoir numerical simulation agent model is as follows:
[0017] Step 2.1, inputting reservoir parameters into a feature extraction module to extract complex spatial features of reservoir parameters;
[0018] Step 2.2, input the extracted features into the feature conversion module, extract the hidden layer features of the reservoir parameters through the global pooling layer, and add the random discarding neuron structure of the random discarding layer to make the proxy model output the uncertainty of the prediction. Finally, based on the replication layer, the reservoir parameter features are converted into parameters with time series characteristics;
[0019] Step 2.3: Input the time series parameters into the time series prediction module, obtain the initial value of the reservoir production data based on the attention mechanism, and then pass it through a layer of activation function to obtain the final production data.
[0020] Furthermore, in step 2, the specific process of training the proxy model is as follows:
[0021] First, set the training parameters of the proxy model, including learning rate, number of training times, batch size, optimizer, and loss function; the loss function is the mean square error between the predicted value and the true value, and the formula is as follows:
[0022] (1);
[0023] In the formula, is the mean square error; is the sample size; For the The true value of samples; Indicates The proxy model prediction value of samples;
[0024] Then, the reservoir numerical simulation data set is divided into a training set and a test set in proportion. The proxy model is trained based on the training set, and the test set is used to verify the prediction effect of the proxy model on the production data.
[0025] Furthermore, the specific process of step 3 is as follows:
[0026] Step 3.1, use the ensemble smoothing optimization algorithm to iteratively adjust the reservoir model parameters, set the parameter set size and the maximum number of iterations; input the updated model parameters into the proxy model, and predict the production data of the updated model;
[0027] Step 3.2: Calculate the data fitting error index to evaluate the model update effect. The formula is as follows:
[0028] (2);
[0029] In the formula, is the data fitting error index; represents observation data; represents the model parameters; represents the mean of the model prior parameters; Represents the proxy model; is transposed; represents the observed data error covariance; represents the covariance of model parameters.
[0030] Furthermore, the specific process of step 4 is as follows:
[0031] Step 4.1: Construct an online learning mechanism. After each optimization iteration, take the reservoir model parameter set as input, use the proxy model for multiple forward predictions, and activate the random discard neuron structure in the proxy model to obtain multiple different prediction values. Calculate the uncertainty of the production data prediction of each reservoir model in the set. The formula is as follows:
[0032] (3);
[0033] In the formula, Based on the proxy model The uncertainty of the predictions made by the reservoir model; Indicates the quantity of production data; Represents the number of forward predictions made by the surrogate model; Represents the proxy model for Reservoir Model The average value of multiple forward forecasts of production data; Represents the proxy model for The first Production data of The next forward prediction value;
[0034] Step 4.2: Sort the uncertainty of each reservoir model prediction in descending order according to the proxy model, and select the top 25% of reservoir model parameters.
[0035] Furthermore, the specific process of step 6 is as follows:
[0036] The final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3 is analyzed, the mean and standard deviation indicators of the reservoir model parameter distribution are calculated, and the uncertainty of the parameter distribution is evaluated. The formula is as follows:
[0037] (4);
[0038] (5);
[0039] In the formula, is the mean value of the final reservoir model parameters; is the standard deviation of the final reservoir model parameters; Indicates the first Parameters corresponding to each reservoir model; Represents the number of reservoir models.
[0040] The beneficial technical effects brought by the present invention are as follows: the present invention establishes a reservoir numerical simulation proxy model based on the Transformer model. In the iterative optimization of reservoir parameters, combined with the online learning mechanism, samples with larger uncertainty in the reservoir model set are selected for numerical simulation, and the proxy model is trained using the latest data. Therefore, the present invention has better applicability and does not need to pre-generate a large amount of numerical simulation data to train the proxy model. Instead, the proxy model is trained using part of the data in the iterative process of the optimization algorithm. This method fully combines the proxy model with the numerical simulation technology, and while obtaining better parameter inversion results, it significantly reduces the computational time of reservoir parameter inversion. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 The present invention is a flow chart of the reservoir parameter inversion modeling method integrating the proxy model and online learning.
[0042] Figure 2 This is a distribution diagram of parameters of the real reservoir model of the present invention.
[0043] Figure 3 This is a structural diagram of the reservoir numerical simulation agent model based on the Transformer model of the present invention.
[0044] Figure 4 This is a structural diagram of the Transformer layer of the present invention.
[0045] Figure 5 This is a comparison effect diagram of the data fitting error objective function of the present invention.
[0046] Figure 6 It is a comparison chart of the fitting effect of the oil production rate data of the production well 1 of the present invention.
[0047] Figure 7 It is a comparison chart of the fitting effect of water production rate data of production well 1 of the present invention.
[0048] Figure 8 This is a mean value diagram of reservoir parameter inversion based on the numerical simulation method of the present invention.
[0049] Fig. 9 This is a standard deviation diagram of reservoir parameter inversion based on the numerical simulation method of the present invention.
[0050] Fig.10 This is a mean value diagram of reservoir parameter inversion based on the proxy model method of the present invention.
[0051] Fig.11 This is a standard deviation diagram of reservoir parameter inversion based on the proxy model method of the present invention. DETAILED DESCRIPTION
[0052] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0053] like Figure 1 As shown, a reservoir parameter inversion modeling method integrating a proxy model and online learning includes the following steps:
[0054] Step 1: Combine the reservoir numerical simulator with prior parameters to obtain reservoir model parameters and simulated production data, and construct a reservoir numerical simulation data set. The specific process is as follows:
[0055] According to the prior parameter samples of the reservoir model, a reservoir numerical simulator is used to perform numerical simulation to obtain reservoir model parameters and corresponding simulated production data; the reservoir model parameters include grid permeability, porosity, relative permeability, etc., and the simulated production data include oil production, water production, pressure data, etc. The reservoir model parameters and simulated production data form a reservoir numerical simulation data set, which is used to train and verify the reservoir numerical simulation proxy model constructed in step 2.
[0056] In the embodiment of the present invention, the number of grids in the reservoir model in the x, y, and z directions is 20, 20, and 1, respectively, for a total of 400 grids, and the sizes of the grids in the x, y, and z directions are 8 m, 8 m, and 4 m, respectively. The model porosity is 0.2, and the net-to-gross ratio is 1. The x-direction permeability distribution of the real model is as follows: Figure 2 As shown, Figure 2In the figure, mD is the unit of permeability, and log indicates that the permeability value in the figure is the value after taking the logarithm. The permeability value in the y direction of the model is the same as the permeability value in the x direction, and the permeability in the z direction is 0.1 times the permeability in the x direction. The model includes 1 water injection well (I1) and 4 production wells (P1-P4). The working system of the injection well is to carry out fixed injection volume, and the injection rate is 100 cubic meters / day. The working system of the production well is to produce at a fixed bottom hole pressure, and the pressure value is 340bar. The reservoir model considers oil and water phases for numerical simulation. The production time is 360 days, which is divided into 36 times for numerical simulation. The reservoir inversion parameter is the permeability in the x direction, and the observation data include the oil production rate and water production rate of the 4 production wells.
[0057] According to the 500 prior parameter samples of the reservoir model, reservoir numerical simulation is carried out to obtain 500 sets of reservoir model parameters and corresponding simulated production data.
[0058] Step 2: construct a reservoir numerical simulation proxy model based on the Transformer model, and train the proxy model in combination with the reservoir numerical simulation data set;
[0059] like Figure 3 As shown, the reservoir numerical simulation agent model includes a feature extraction module, a feature conversion module, and a time series prediction module; wherein the feature extraction module includes structures such as a convolution layer, a downsampling layer, and an activation function; in an embodiment of the present invention, the structure of the convolutional neural network model may include three groups of sequentially connected convolution layers, downsampling layers, and activation functions (i.e., corresponding to Figure 3 middle ). The feature conversion module includes a global pooling layer, a random drop layer, and a replication layer. The time series prediction module includes multiple Transformer layers. The structure of each Transformer layer is as follows: Figure 4 As shown, it includes a multi-head attention layer, layer normalization, and a linear layer, and considers the addition operation of cross-layer parameters, which can improve the training effect of the proxy model. In this embodiment of the present invention, the number of Transformer layers is 4 (corresponding to Figure 3 middle ).
[0060] The specific working process of the reservoir numerical simulation agent model is as follows:
[0061] Step 2.1, inputting reservoir parameters into a feature extraction module to extract complex spatial features of reservoir parameters;
[0062] Step 2.2, input the extracted features into the feature conversion module, extract the hidden layer features of the reservoir parameters through the global pooling layer, and add the random dropout neuron structure (Dropout) of the random dropout layer. The dropout ratio is set to 0.2, so that the proxy model can output the uncertainty of the prediction. Finally, based on the replication layer, the reservoir parameter features are converted into parameters with time series characteristics;
[0063] Step 2.3: Input the time series parameters into the time series prediction module, and based on the attention mechanism, obtain the initial value of the reservoir production data. Then, after a layer of activation function, the final production data is obtained, thereby realizing the prediction of the reservoir production data.
[0064] The specific process of training the proxy model is:
[0065] First, set the training parameters of the proxy model, including learning rate, number of training times, batch size, optimizer, and loss function. The loss function is the mean square error between the predicted value and the true value, and the formula is as follows:
[0066] (1);
[0067] In the formula, is the mean square error; is the sample size; For the The true value of samples; Indicates The surrogate model predicts the value of samples.
[0068] Then, the reservoir numerical simulation data set is divided into a training set and a test set according to the ratio of 8:2. The proxy model is trained based on the training set, and the test set is used to verify the prediction effect of the proxy model on the production data.
[0069] In the embodiment of the present invention, the learning rate is 0.001, the number of training times is 500, the batch size is 16, and the optimizer is the Adam algorithm. The mean square error on the training set can reach 0.002, and the mean square error on the test set can reach 0.01, showing that the proxy model has good prediction effects in both the training set and the test set.
[0070] Step 3: Use the optimization algorithm to iteratively adjust the reservoir model parameter set, and then use the proxy model to evaluate the data fitting effect of the updated model; the specific process is as follows:
[0071] Step 3.1: Use the ensemble smoothing optimization algorithm to iteratively adjust the reservoir model parameters, set the parameter set size to 200, and set the maximum number of iterations to 10. Input the updated model parameters into the proxy model and predict the production data of the updated model.
[0072] Step 3.2: Calculate the data fitting error index to evaluate the model update effect. The formula is as follows:
[0073] (2);
[0074] In the formula, is the data fitting error index; represents observation data; represents the model parameters; represents the mean of the model prior parameters; Represents the proxy model; is transposed; represents the observed data error covariance; represents the covariance of model parameters.
[0075] Step 4: In the iteration of the optimization algorithm, an online learning mechanism is constructed to determine the uncertainty of the proxy model prediction based on the random discarding neuron structure, and select the reservoir model parameters with large uncertainty; the specific process is as follows:
[0076] Step 4.1: Construct an online learning mechanism. After each optimization iteration, take the reservoir model parameter set as input, use the proxy model for multiple forward predictions, and activate the random dropout neuron structure (Dropout) in the proxy model to obtain multiple different prediction values. Calculate the uncertainty of the production data prediction of each reservoir model in the set. The formula is as follows:
[0077] (3);
[0078] In the formula, Based on the proxy model The uncertainty of the predictions made by the reservoir model; Indicates the quantity of production data; Represents the number of forward predictions made by the surrogate model; Represents the proxy model for Reservoir Model The average value of multiple forward forecasts of production data; Represents the proxy model for The first Production data of The forward prediction value.
[0079] Step 4.2: Sort the uncertainty of each reservoir model prediction in descending order according to the proxy model, and select the top 25% of reservoir model parameters.
[0080] Step 5: Perform numerical simulation on the selected reservoir model parameters, obtain new simulated production data, construct a new reservoir numerical simulation data set, and retrain the proxy model using all current data sets;
[0081] Step 6: According to the final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3, the distribution of actual reservoir parameters is determined to complete the reservoir parameter inversion process.
[0082] The final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3 is analyzed, the mean and standard deviation indicators of the reservoir model parameter distribution are calculated, and the uncertainty of the parameter distribution is evaluated. The formula is as follows:
[0083] (4);
[0084] (5);
[0085] In the formula, is the mean value of the final reservoir model parameters; is the standard deviation of the final reservoir model parameters; Indicates the first Parameters corresponding to each reservoir model; Represents the number of reservoir models.
[0086] The method based on numerical simulation requires a total of 2000 numerical simulations. The method based on surrogate model uses the surrogate model to evaluate the prediction effect after each iteration and selects 25% of the samples for simulation, requiring a total of 500 numerical simulations, which is a 75% reduction in the number of numerical simulations.
[0087] Figure 5 It is the distribution of the objective function of the data fitting error, which includes the initial set (abbreviated as initial in the figure), the final set based on the numerical simulation method (abbreviated as digital model in the figure), and the final set based on the proxy model of the present invention (abbreviated as the proposed method in the figure). It can be seen that the objective function values obtained based on the numerical simulation method and the proxy model method are relatively low, indicating that the data fitting error is greatly reduced compared with the initial set. Comparing the numerical simulation method with the proxy model method, it can be found that the data fitting errors obtained by the two methods are very close, which verifies the reliability of the proxy model method.
[0088] Figure 6 , Figure 7The comparison of the fitting effect of the observed data (abbreviated as observation in the figure) of the oil production rate and water production rate of production well 1. The curves in the figure represent the initial set (abbreviated as initial in the figure), the final set obtained by the numerical simulation method (abbreviated as numerical simulation in the figure), and the mean of the final set obtained by the proxy model method (abbreviated as proposed in the figure). It can be seen that the error between the simulated data of the initial model set and the observed data is large, while the optimized data almost coincides with the observed data, showing a good fitting effect. At the same time, the data fitting effect based on the numerical simulation method is close to the data fitting effect based on the proxy model.
[0089] Figure 8 , Fig. 9 , Fig.10 , Fig.11 The figure shows the comparison of the mean and standard deviation of the inversion model parameters obtained based on the numerical simulation method and the proxy model method. It can be seen that the mean of the ensemble model parameters obtained by the two methods is close to the distribution of the true model parameters, especially at the production well location. From the standard deviation diagram of the final model parameters, it can be seen that the standard deviation at the production well location is small, while the standard deviation at other locations is large, which shows the uncertainty of the inversion parameters, that is, the larger the standard deviation value, the greater the uncertainty of the parameter. The inversion parameter results obtained based on the numerical simulation method and the proxy model method are also relatively close.
[0090] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A reservoir parameter inversion modeling method integrating a proxy model and online learning, characterized in that: The steps include: Step 1: Combine the reservoir numerical simulator with prior parameters to obtain reservoir model parameters and simulated production data, and construct a reservoir numerical simulation data set; Step 2: construct a reservoir numerical simulation proxy model based on the Transformer model, and train the proxy model in combination with the reservoir numerical simulation data set; Step 3: Use the optimization algorithm to iteratively adjust the reservoir model parameter set, and then use the proxy model to evaluate the data fitting effect of the updated model; Step 4: In the iteration of the optimization algorithm, an online learning mechanism is constructed to determine the uncertainty of the proxy model prediction based on the random discarding neuron structure and select the reservoir model parameters with large uncertainty; Step 5: Perform numerical simulation on the selected reservoir model parameters, obtain new simulated production data, construct a new reservoir numerical simulation data set, and retrain the proxy model using all current data sets; Step 6: According to the final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3, the distribution of actual reservoir parameters is determined to complete the reservoir parameter inversion process.
2. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 1, characterized in that: In the step 1, numerical simulation is performed using a reservoir numerical simulator based on a priori parameter samples of the reservoir model to obtain reservoir model parameters and corresponding simulated production data; wherein the reservoir model parameters include grid permeability, porosity, and relative permeability, and the simulated production data include oil production, water production, and pressure data.
3. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 1, characterized in that: In step 2, the reservoir numerical simulation proxy model includes a feature extraction module, a feature conversion module, and a time series prediction module; wherein the feature extraction module includes a convolution layer, a downsampling layer, and an activation function; the feature conversion module includes a global pooling layer, a random drop layer, and a replication layer; the time series prediction module includes multiple Transformer layers, each Transformer layer includes a multi-head attention layer, a layer normalization, a linear layer, and also includes an addition operation of cross-layer parameters; The specific working process of the reservoir numerical simulation agent model is as follows: Step 2.1, inputting reservoir parameters into a feature extraction module to extract complex spatial features of reservoir parameters; Step 2.2, input the extracted features into the feature conversion module, extract the hidden layer features of the reservoir parameters through the global pooling layer, and add the random discarding neuron structure of the random discarding layer to make the proxy model output the uncertainty of the prediction. Finally, based on the replication layer, the reservoir parameter features are converted into parameters with time series characteristics; Step 2.3: Input the time series parameters into the time series prediction module, obtain the initial value of the reservoir production data based on the attention mechanism, and then pass it through a layer of activation function to obtain the final production data.
4. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 1, characterized in that: In step 2, the specific process of training the proxy model is as follows: First, set the training parameters of the proxy model, including learning rate, number of training times, batch size, optimizer, and loss function; the loss function is the mean square error between the predicted value and the true value, and the formula is as follows: (1)； In the formula, is the mean square error; is the sample size; For the The true value of samples; Indicates The proxy model prediction value of samples; Then, the reservoir numerical simulation data set is divided into a training set and a test set in proportion. The proxy model is trained based on the training set, and the test set is used to verify the prediction effect of the proxy model on the production data.
5. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 4 is characterized in that: The specific process of step 3 is as follows: Step 3.1, use the ensemble smoothing optimization algorithm to iteratively adjust the reservoir model parameters, set the parameter set size and the maximum number of iterations; input the updated model parameters into the proxy model, and predict the production data of the updated model; Step 3.2: Calculate the data fitting error index to evaluate the model update effect. The formula is as follows: (2)； In the formula, is the data fitting error index; represents observation data; represents the model parameters; represents the mean of the model prior parameters; Represents the proxy model; is transposed; represents the observed data error covariance; represents the covariance of model parameters.
6. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 5 is characterized in that: The specific process of step 4 is as follows: Step 4.1: Construct an online learning mechanism. After each optimization iteration, take the reservoir model parameter set as input, use the proxy model for multiple forward predictions, and activate the random discard neuron structure in the proxy model to obtain multiple different prediction values. Calculate the uncertainty of the production data prediction of each reservoir model in the set. The formula is as follows: (3)； In the formula, Based on the proxy model The uncertainty of the predictions made by the reservoir model; Indicates the quantity of production data; Represents the number of forward predictions made by the surrogate model; Represents the proxy model for Reservoir Model The average value of multiple forward forecasts of production data; Represents the proxy model for The first Production data of The next forward prediction value; Step 4.2: Sort the uncertainty of each reservoir model prediction in descending order according to the proxy model, and select the top 25% of reservoir model parameters.
7. The reservoir parameter inversion modeling method integrating agent model and online learning according to claim 6, characterized in that: The specific process of step 6 is as follows: The final reservoir model parameter set obtained after iterative solution of the optimization algorithm in step 3 is analyzed, the mean and standard deviation indicators of the reservoir model parameter distribution are calculated, and the uncertainty of the parameter distribution is evaluated. The formula is as follows: (4)； (5)； In the formula, is the mean value of the final reservoir model parameters; is the standard deviation of the final reservoir model parameters; Indicates the first Parameters corresponding to each reservoir model; Represents the number of reservoir models.