A geological modeling method based on neural network and dynamic data assimilation

Abstract

The present invention relates to a geological modeling method based on a neural network and dynamic data assimilation. The method uses a generator neural network to establish an initial geological model according to an initial input vector. Secondly, production dynamic fitting data corresponding to the initial geological model is obtained according to a preset physical constraint proxy network. Based on the ES-MDA method, data assimilation operation is performed according to the fitted data and the actually measured data, the input vector of the generator neural network is adjusted according to the data error, and the geological model is iteratively updated to finally obtain an optimal geological model. In the present invention, the high-dimensional geological model is represented by a one-dimensional random vector through the generator neural network, and the input vector is adjusted instead of adjusting the model parameters, reducing the calculation time and cost; data assimilation of the input vector is realized through the ES-MDA method, and the generator network generates a geological model that meets the dynamic data based on the assimilated input vector, improving the modeling efficiency and model accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of geological modeling, and particularly to a geological modeling method based on neural network and dynamic data assimilation. Background Art

[0002] In the field of reservoir engineering, geological modeling is of extremely important significance. Geological modeling helps researchers better understand and predict the geological structure, lithology distribution, reservoir characteristics, etc. of oil and gas reservoirs by digitally describing and simulating underground geological bodies, providing key support for the exploration, development, and production of reservoir engineering. Geological modeling provides a reliable scientific basis for oil and gas exploration and development, contributing to improving the success rate of exploration and development and resource utilization efficiency.

[0003] Existing geological models are usually established using methods such as stochastic simulation modeling, physical modeling, and numerical simulation. The stochastic simulation method constructs multiple groups of geological models by randomly generating the values of geological parameters and then uses physical simulation or numerical simulation methods for simulation prediction. This method has the problems of weak dependence on actual production data, lack of real-time performance and accuracy. The physical modeling method constructs physical models of geological models, such as hydrodynamic models and in-situ stress models, based on geological theories and experimental data. This method usually requires a large amount of static geological data input and is difficult to adjust and update in real time. In addition, in classical geostatistics, manually adjusting geological models is time-consuming and laborious and highly dependent on expert experience, and the non-Gaussian characteristics of geological models bring difficulties to the automatic update of model parameters.

[0004] Therefore, it is necessary to propose a geological modeling method based on neural network and dynamic data assimilation, which can take into account the uncertainty of actual dynamic data, quickly generate geological models, and optimize model parameters in combination with measured data to improve the reliability and accuracy of geological modeling. Summary of the Invention

[0005] In view of this, the present invention provides a geological modeling method based on neural network and dynamic data assimilation to solve the technical problems in the prior art that manual adjustment of geological model parameters is time-consuming and laborious and highly dependent on expert experience, and the geological model cannot be adjusted and optimized in real time.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a geological modeling method based on neural network and dynamic data assimilation, including:

[0008] Establishing an initial geological model using a generator neural network according to an input vector;

[0009] Input the initial geological model into a preset physical constraint surrogate network to obtain the production performance fitting data of the initial geological model;

[0010] Based on the measured production performance data and the production performance fitting data, use the ES-MDA method for data assimilation to adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, obtain the optimal geological model.

[0011] Furthermore, the physical constraint surrogate network model is constructed based on a CNN network, and the loss function is:

[0012] L ph =‖PDEs[sur(m,c),m,c]‖ p

[0013] where ‖·‖ p represents Lp regularization, m represents the model parameters of the geological model, c represents the simulation parameters, sur() represents the production performance fitting data obtained according to the geological model and the simulation parameters, and PDEs represents the preset partial differential equation, which is used to express the preset physical constraint relationship satisfied by the production performance fitting data output by the preset physical constraint surrogate network, the geological model, and the simulation parameters.

[0014] Furthermore, when the geological model is a two-dimensional grid model, the expression of the preset partial differential equation is:

[0015]

[0016] where (x,y) represents the coordinates, which are the grid positions in the geological model, t represents the time, k(x,y) represents the model parameters at the position (x,y), μ and C are the simulation parameters, and A is a constant.

[0017] Furthermore, based on the measured production performance data and the production performance fitting data, using the ES-MDA method for data assimilation includes:

[0018] Calculate the production performance prediction data at the current time according to the geological model, the simulation parameters, and the production performance fitting data at the previous time;

[0019] Based on the production performance prediction data at the current time and the measured production performance data at the current time, calculate the data error value;

[0020] Determine the production performance analysis data at the current time according to the data error value, the production performance prediction data at the current time, and the production performance data measured at the current time; and calculate the production performance prediction data at the next time based on the production performance analysis data at the current time.

[0021] Further, production dynamic analysis data at the current moment is determined based on the data error value, production dynamic prediction data at the current moment, and production dynamic measured data at the current moment, including:

[0022] Production dynamic analysis data at the current moment Production dynamic prediction data x at the current moment t And production dynamic measured data Y at the current moment t The relational expression is:

[0023]

[0024]

[0025] where H() represents an observation operator for aligning the dimensions of production dynamic data and production dynamic measured data; P k represents the error covariance matrix of the prediction data at the current moment, which is calculated by minimizing the error between the production dynamic prediction data and the production dynamic measured data; R represents the first Gaussian white noise; α represents the inflation coefficient.

[0026] Further, based on the production dynamic analysis data at the current moment, the production dynamic prediction data for the next moment is calculated, including:

[0027] Production dynamic prediction data x for the next moment t+1 And the production dynamic analysis data at the current moment The relational expression is:

[0028]

[0029] where M() represents a non-linear model operator, and w k represents the second Gaussian white noise.

[0030] Further, the production dynamic fitting data includes wellhead pressure, oil production rate, and water production rate; the model parameters include porosity, permeability, and water saturation; the simulation parameters include fluid viscosity, compressibility of the rock and the internal fluid.

[0031] In a second aspect, the present invention also provides a geological modeling device based on a neural network and dynamic data assimilation, including:

[0032] A model generation module for establishing an initial geological model using a generator neural network according to an input vector;

[0033] A dynamic data generation module for inputting the initial geological model into a preset physical constraint proxy network to obtain production dynamic fitting data of the initial geological model;

[0034] An optimization module, which is used to perform data assimilation based on the measured production dynamic data and the production dynamic fitting data by using the ES-MDA method, adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, an optimal geological model is obtained.

[0035] In a third aspect, the present invention also provides an electronic device, including a processor and a memory, where a computer program is stored on the memory, and when the computer program is executed by the processor, the geological modeling method based on neural network and dynamic data assimilation according to any one of the above technical solutions is implemented.

[0036] In a fourth aspect, the present invention also provides a computer-readable storage medium for storing computer-readable programs or instructions, and when the programs or instructions are executed by a processor, the steps in the geological modeling method based on neural network and dynamic data assimilation in any one of the above implementation manners can be implemented.

[0037] The present invention provides a geological modeling method based on neural network and dynamic data assimilation. First, an initial geological model is established by using a generator neural network according to an initial input vector; secondly, production dynamic fitting data corresponding to the initial geological model is obtained according to a preset physical constraint proxy network; finally, based on the ES-MDA method, data assimilation operations are performed according to the fitted data and the actually measured data, the input vector of the generator neural network is adjusted according to the data error, and the geological model is iteratively updated to finally obtain an optimal geological model. The method of the present invention uses the generator of the GANs network to represent a high-dimensional model with a low-dimensional random vector, parameterizes the geological model into a group of one-dimensional random vectors, and adjusts the input vector instead of adjusting the model parameters, thereby significantly reducing the calculation time and cost; using a preset physical constraint proxy network to determine production dynamic fitting data according to the model can more accurately simulate the state change of the geological model in actual production and improve the reliability and fidelity of the model; since the input vector is a Gaussian distribution and the dimension is much lower than the dimension of the geological model, the ES-MDA method can be used to perform data assimilation on the input vector to optimize and correct the model parameters. The present invention can not only automatically fit reservoir production data, but also take into account the uncertainty of the data, improve the reliability and accuracy of the model, and provide a more accurate prediction and decision-making basis for actual production. Description of the Drawings

[0038] Figure 1 It is a schematic flowchart of the geological modeling method based on neural network and dynamic data assimilation provided by the present invention;

[0039] Figure 2 It is a training schematic diagram of an embodiment of the generator network model provided by the present invention;

[0040] Figure 3 Schematic structural diagram of an embodiment of the preset physical constraint proxy network provided by the present invention;

[0041] Figure 4 Overall schematic diagram of an embodiment of the modeling process provided by the present invention;

[0042] Figure 5 Schematic structural diagram of an embodiment of the geological modeling device based on neural network and dynamic data assimilation provided by the present invention;

[0043] Figure 6 Schematic structural diagram of an embodiment of the electronic device provided by the present invention. Detailed implementation manners

[0044] The preferred embodiments of the present invention will be specifically described below with reference to the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0045] Before introducing the embodiments of the present invention, relevant terms and inventive concepts of this application will be explained.

[0046] GANs neural network: Generative Adversarial Networks (GAN) is a deep learning model composed of two parts: a generator and a discriminator. The generative adversarial network usually takes a random vector that satisfies the Gaussian distribution as input. Its goal is to generate fake data that is as close as possible to the real data, while the goal of the discriminator is to distinguish whether the input data comes from the real data set or is generated by the generator. Through this adversarial process, the generator learns how to generate more and more realistic data. In this solution, the generator of the GANs network represents the high-dimensional model through a low-dimensional random vector, playing a role in dimensionality reduction. In this way, it is not necessary to directly modify the geological model, but to indirectly modify the model by modifying the input vector, thereby significantly reducing the calculation time and cost.

[0047] ES-MDA: The Ensemble Smoother with Multiple Data Assimilation (ES-MDA) method is developed based on the Ensemble Kalman Filter (EnKF) and has lower computational costs, better data fitting, and parameter estimation effects, making it the main application method currently. Emerick (2016) analyzed the production and seismic data assimilation performance using ES-MDA and verified the excellent performance of ES-MDA in data assimilation. The ES-MDA method is only applicable to Gaussian fields (Emerick, 2013), while generative adversarial networks usually take random vectors that satisfy the Gaussian distribution as inputs and have an innate advantage in combination with ES-MDA.

[0048] Ensemble Kalman Filter (EnKF): An efficient data assimilation method used to fuse observational data into a numerical model to improve the model's state estimation. In the field of reservoir management, EnKF represents the model's uncertainty by using a series of model predictions (i.e., ensemble members) and uses actual production data to update these predictions, thereby reducing the uncertainty of the predictions and improving the accuracy of the model predictions.

[0049] The present invention provides a geological modeling method, device, electronic device, and computer-readable storage device based on neural networks and dynamic data assimilation, which will be described separately below.

[0050] Combined Figure 1 As shown, a specific embodiment of the present invention discloses a geological modeling method based on neural networks and dynamic data assimilation, including:

[0051] Step S101: Use a generator neural network to establish an initial geological model based on an input vector;

[0052] Step S102: Input the initial geological model into a preset physical constraint proxy network to obtain production dynamic fitting data of the initial geological model;

[0053] Step S103: Based on the actual production dynamic data and the production dynamic fitting data, use the ES-MDA method for data assimilation, adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, obtain an optimal geological model.

[0054] Compared with the prior art, the method of this embodiment first uses a generator neural network to establish an initial geological model according to an initial input vector; secondly, obtains production dynamic fitting data corresponding to the initial geological model according to a preset physical constraint proxy network; finally, based on the ES-MDA method, performs data assimilation operations according to the fitted data and the actually measured data, adjusts the input vector of the generator neural network according to the data error, and iteratively updates the geological model to finally obtain an optimal geological model. The method of this embodiment uses the powerful data generation ability of GANs to create a high-quality reservoir model parameter set, which can effectively reduce the number of model runs required in the collection assimilation process, thereby significantly reducing the calculation time and cost; uses a preset physical constraint proxy network to determine production dynamic fitting data according to the model, can more accurately simulate the state change of the geological model in actual production, and improves the reliability and fidelity of the model; finally, assimilates the actual observed data into the model through the ES-MDA method to optimize and correct the model parameters. Compared with the data-driven dynamic data fitting method, the method combining data assimilation and deep learning has stronger operability and robustness. The method of this embodiment can not only automatically fit reservoir production data, but also take into account the uncertainty of the data, improve the reliability and accuracy of the model, and provide a more accurate prediction and decision-making basis for actual production.

[0055] It should be noted that, in order to train the generator neural network (simulator), a discriminator is also required in practice. The discriminator is only used to train the generator. In our workflow, only the trained generator is used, and the discriminator is not used. As Figure 2 shown, Figure 2 shows the training process of the GANs neural network (taking the output model as 64x64 as an example). The loss function for training is:

[0056]

[0057] where L adv (G n , D n ) is the conditional adversarial network loss. The training objective of the generator is to maximize the gap between the discriminator for the generated target and the real sample under preset conditions, and the training objective of the discriminator is to minimize the distance between the real sample and the generated sample.

[0058] After training is completed, the core function of the generator is to be able to obtain a geological model according to the input random vector. Using the powerful data generation ability of GANs to create a high-quality reservoir model parameter set can reduce the number of model runs required in the collection assimilation process, thereby significantly reducing the calculation time and cost.

[0059] As a preferred embodiment, the production dynamic fitting data obtained by the preset physical proxy network model includes wellhead pressure, oil production, and water production; the model parameters include porosity, permeability, and water saturation; the simulation parameters include fluid viscosity, compressibility of rock and internal fluid.

[0060] As a preferred embodiment, the physical constraint proxy network model is constructed based on a CNN network, and the loss function is:

[0061] L ph =‖PDEs[sur(m,c),m,c]‖ p (2)

[0062] where ‖·‖ p represents Lp regularization, p = 1 represents taking the absolute value, p = 2 represents taking the square; m represents the model parameters of the geological model, c represents the simulation parameters, sur() represents the production dynamic fitting data obtained according to the geological model and simulation parameters, and PDEs represents the preset partial differential equation, which is used to express the preset physical constraint relationship satisfied by the production dynamic fitting data output by the preset physical constraint proxy network, the geological model, and the simulation parameters.

[0063] In a specific embodiment, optimizing the loss function of formula (2) above is the process of training the physical constraint proxy network. Taking the pressure map as the production dynamic data as an example, the loss function reflects the gap between the output pressure map and the measured real pressure map. Optimizing the loss function to reduce the difference between the two is the process of training the network.

[0064] In a specific embodiment, as Figure 3 shown, Figure 3 shows a possible structure of the physical constraint proxy network model. Taking 64x64 input and output as an example, the role of the physical constraint proxy network is to convert the geological model output by the generator into production dynamic data. Here, the pressure map is used as the production dynamic data. The training process needs to be based on a dataset. Taking the grid structure diagram as an example, the inputs include the geological model, namely: permeability model, initial pressure model, boundary pressure model, and time step, etc. After passing through the grid, the output is the pressure map. The fitting dynamic data, geological model, and simulation parameters satisfy the preset physical laws through partial differential equations, and the loss function of the model is optimized, thereby realizing the training of the neural network under physical constraints.

[0065] As a preferred embodiment, when the geological model is a two-dimensional grid model, the expression of the preset partial differential equation is:

[0066]

[0067] Among them, (x, y) represents coordinates, which are the grid positions in the geological model, marking the grid positions in the 64x64 grid model, with the unit of meter, and can be understood as the distance from the upper left corner of the 64x64 model; t represents the time, in days (the production dynamic data changes every day); k(x, y) represents the permeability at the position (x, y), μ is the fluid viscosity, C is the compressibility of the rock and the fluid inside it, and A is a constant unit conversion factor. μ and C are the simulation parameters c in formula (2), and k(x, y) is m in formula (2).

[0068] As a preferred embodiment, based on the measured production dynamic data and the production dynamic fitting data, the ES-MDA method is used for data assimilation, including:

[0069] Calculate the production dynamic prediction data at the current time according to the geological model, simulation parameters and the production dynamic fitting data at the previous time;

[0070] Based on the production dynamic prediction data at the current time and the measured production dynamic data at the current time, calculate the data error value;

[0071] Determine the production dynamic analysis data at the current time according to the data error value, the production dynamic prediction data at the current time and the production dynamic measured data at the current time; and calculate the production dynamic prediction data at the next time based on the production dynamic analysis data at the current time.

[0072] It should be noted that the idea of the above processing process is: predict the production dynamic data at the current time through the production dynamic fitting data at the previous time. Since there must be errors between the prediction and the actual situation, it is necessary to measure the current data for auxiliary analysis. However, there are also errors in the measurement, so the analysis data is obtained by combining the current measurement data and the predicted production dynamic data; calculate the production dynamic prediction data at the next time based on the analysis data to improve the reliability and accuracy of the model prediction.

[0073] As a preferred embodiment, determining the production dynamic analysis data at the current time according to the data error value, the production dynamic prediction data at the current time and the production dynamic measured data at the current time includes:

[0074] The production dynamic analysis data at the current time The production dynamic prediction data x at the current time k and the production dynamic measured data Y at the current time k The relational expression is:

[0075]

[0076]

[0077] Among them, H() represents an observation operator, which is used to align the dimensions of production dynamic data and measured production dynamic data; P k represents the error covariance matrix of the predicted data at the current moment, which is calculated by minimizing the error between the predicted production dynamic data and the measured production dynamic data; R represents the first Gaussian white noise; α represents the inflation coefficient.

[0078] The observation operator h() can be understood as an identity transformation. There may be problems of inconsistent units or dimensions between the observed value and the predicted value. By transforming the data through the observation operator, data alignment can be achieved. By introducing the inflation coefficient α, the observation error is updated in each cycle to optimize the effect of data assimilation. Among them, N is the number of cycles.

[0079] As a preferred embodiment, calculating the predicted production dynamic data for the next moment based on the production dynamic analysis data at the current moment includes:

[0080] The predicted production dynamic data x for the next moment t+1 and the production dynamic analysis data at the current moment The relational expression is:

[0081]

[0082] Among them, M() represents a non-linear model operator, w k represents the second Gaussian white noise.

[0083] It should be noted that the first Gaussian white noise R and the second Gaussian white noise w k can be Gaussian white noises with different variance values and expected values.

[0084] ES-MDA data assimilation is actually a recursive formula. From Z k+1 =Z k +e k , e k is the error. k can be understood as the number of iteration rounds. By measuring the production data every day and combining the predicted values of the neural network, Z can be adjusted to obtain the most suitable input vector.

[0085] Combined with Figure 4 to further elaborate on the above technical solution, such as Figure 4As shown in the figure, the generator of the GANs network is selected to generate the initial model. The GANs modeling parameterizes the geological model into a set of one-dimensional random vectors, representing the high-dimensional model with low-dimensional random vectors, which plays a role in dimensionality reduction. In this way, instead of directly modifying the geological model, the model is indirectly modified by modifying the input vector, thus significantly reducing the calculation time and cost, and having the inherent advantage of assimilating production dynamic data. Assume that the initial random vector is Z and the iteration number k = 0. The random vector passes through the generator neural network to obtain the initial geological model.

[0086] After the initial geological model passes through the physical constraint proxy network, the production dynamic fitting data (such as cumulative oil production, wellhead pressure, etc.) of the initial geological model is obtained.

[0087] The ES-MDA method is used to assimilate the fitting production dynamic data with the actually measured production dynamic data. Specifically, based on the production dynamic prediction data and the currently measured production dynamic actual data at the current moment, the data error value is calculated, and the production dynamic analysis data at the current moment is determined according to the data error value, the production dynamic prediction data at the current moment, and the production dynamic actual data at the current moment; and based on the production dynamic analysis data at the current moment, the input vector Z is adjusted, and the updated Z is re-input into the generator neural network to obtain the updated geological model.

[0088] This process is continuously repeated to perform data assimilation on the input vector until the error between the simulated production dynamics and the observed production dynamics is within a given range. The finally obtained geological model matches the measured production dynamics, thus realizing the geological modeling process under the constraint of production dynamics.

[0089] In order to better implement the geological modeling method based on neural network and dynamic data assimilation in the embodiments of the present invention, on the basis of the geological modeling method based on neural network and dynamic data assimilation, correspondingly, please refer to Figure 5 , Figure 5 which is a schematic structural diagram of an embodiment of the geological modeling device based on neural network and dynamic data assimilation provided by the present invention. A geological modeling device 500 based on neural network and dynamic data assimilation provided by an embodiment of the present invention includes:

[0090] A model generation module 501, configured to establish an initial geological model according to an input vector by using a generator neural network;

[0091] A dynamic data generation module 502, configured to input the initial geological model into a preset physical constraint proxy network to obtain the production dynamic fitting data of the initial geological model;

[0092] An optimization module 503 is configured to perform data assimilation using the ES-MDA method based on the measured production dynamic data and the production dynamic fitting data, adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, an optimal geological model is obtained.

[0093] It should be noted here that: the corresponding device 700 provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be referred to the corresponding content in the above method embodiments, and will not be elaborated here.

[0094] As Figure 6 shown, for the above-mentioned geological modeling method based on neural network and dynamic data assimilation, the present invention also correspondingly provides an electronic device 600, which can be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer, and a server. The electronic device includes a processor 601, a memory 602, and a display 603.

[0095] The memory 602 may be an internal storage unit of the computer device in some embodiments, such as the hard disk or memory of the computer device. The memory 602 may also be an external storage device of the computer device in other embodiments, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the computer device. Further, the memory 602 may include both the internal storage unit and the external storage device of the computer device. The memory 602 is used to store the application software installed in the computer device and various types of data, such as the program code installed in the computer device. The memory 602 may also be used to temporarily store the data that has been output or will be output. In one embodiment, a program 604 of a geological modeling method based on neural network and dynamic data assimilation is stored on the memory 602, and the program 604 of the geological modeling method based on neural network and dynamic data assimilation can be executed by the processor 601, so as to implement the geological modeling method based on neural network and dynamic data assimilation in various embodiments of the present invention.

[0096] The processor 601 may be a central processing unit (CPU), a microprocessor, or other data processing chips in some embodiments, and is used to run the program code stored in the memory 602 or process data, such as executing a program of a geological modeling method based on neural network and dynamic data assimilation.

[0097] The display 603 may be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch screen, etc. in some embodiments. The display 603 is used to display information of the computer device and to display a visual user interface. Components 601-603 of the computer device communicate with each other via a system bus.

[0098] This embodiment also provides a computer-readable storage medium, on which a geological modeling program based on neural network and dynamic data assimilation is stored. When the geological modeling program based on neural network and dynamic data assimilation is executed by a processor, the steps in the above embodiment can be implemented.

[0099] The present invention provides a geological modeling method based on neural network and dynamic data assimilation. The method of this application and the method of this embodiment first use a generator neural network to establish an initial geological model according to an initial input vector; secondly, obtain production dynamic fitting data corresponding to the initial geological model according to a preset physical constraint proxy network; finally, based on the ES-MDA method, perform data assimilation operations according to the fitted data and the actually measured data, adjust the input vector of the generator neural network according to the data error, and iteratively update the geological model to finally obtain an optimal geological model.

[0100] The present invention can not only automatically fit reservoir production data, but also take into account the uncertainty of the data, improve the reliability and accuracy of the model, and provide a more accurate prediction and decision-making basis for actual production.

[0101] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A geological modeling method based on neural network and dynamic data assimilation, characterized in that: include: Using a generator neural network to build an initial geological model based on the input vector; The initial geological model is input into a preset physical constraint proxy network to obtain production dynamic fitting data of the initial geological model; the physical constraint proxy network model is constructed based on a CNN network, and the loss function is: L ph =||PDEs[on(m,c),m,c]|| p Among them, ||·|| p represents Lp regularization, m represents the model parameters of the geological model, c represents the simulation parameters, sur() represents the production dynamic fitting data obtained according to the geological model and the simulation parameters, and PDEs represents the preset partial differential equation, which is used to express the preset physical constraint relationship satisfied by the production dynamic fitting data output by the preset physical constraint agent network and the geological model and the simulation parameters; Based on the production dynamic measured data and the production dynamic fitting data, the ES-MDA method is used to perform data assimilation, adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, the optimal geological model is obtained.

2. The geological modeling method based on neural network and dynamic data assimilation according to claim 1, characterized in that: When the geological model is a two-dimensional grid model, the expression of the preset partial differential equation is: Among them, (x, y) represents the coordinates, is the grid position in the geological model, t represents the time, k(x, y) represents the model parameters at the (x, y) position, μ and C are simulation parameters, and A is a constant.

3. The geological modeling method based on neural network and dynamic data assimilation according to claim 1, characterized in that: Based on the production dynamic measured data and the production dynamic fitting data, the ES-MDA method is used to perform data assimilation, including: Calculate the production performance prediction data at the current moment according to the geological model, simulation parameters and the production performance fitting data at the previous moment; Calculate the data error value based on the production dynamics prediction data at the current moment and the production dynamics measured data at the current moment; The production dynamics analysis data at the current moment is determined according to the data error value, the production dynamics prediction data at the current moment and the production dynamics measured data at the current moment; and the production dynamics prediction data at the next moment is calculated based on the production dynamics analysis data at the current moment.

4. The geological modeling method based on neural network and dynamic data assimilation according to claim 3 is characterized in that: Determining the production dynamic analysis data at the current moment according to the data error value, the production dynamic prediction data at the current moment, and the production dynamic measured data at the current moment, including: Current production dynamics analysis data Current production dynamic forecast data x t and the current production dynamic measured data Y t The relationship is: Wherein, H() represents the observation operator, which is used to align the dimensions of production dynamic data with production dynamic measured data; T represents matrix transposition; P k It represents the error covariance matrix of the predicted data at the current moment, which is calculated by minimizing the error between the production dynamics predicted data and the production dynamics measured data; R represents the first Gaussian white noise; α represents the expansion coefficient.

5. The geological modeling method based on neural network and dynamic data assimilation according to claim 3 is characterized in that: Calculate the production dynamics forecast data for the next moment based on the current production dynamics analysis data, including: The production dynamic forecast data x at the next moment t+1 Dynamic analysis data of production at the current moment The relationship is: Among them, M() represents the nonlinear model operator, w k represents the second Gaussian white noise.

6. The geological modeling method based on neural network and dynamic data assimilation according to claim 1, characterized in that: The production dynamic fitting data includes wellhead pressure, oil production, and water production; the model parameters include porosity, permeability, and water saturation; and the simulation parameters include fluid viscosity, rock, and internal fluid compression coefficient.

7. A geological modeling device based on neural network and dynamic data assimilation, characterized in that: include: A model generation module for establishing an initial geological model based on an input vector using a generator neural network; The dynamic data generation module is used to input the initial geological model into a preset physical constraint proxy network to obtain the production dynamic fitting data of the initial geological model; the physical constraint proxy network model is constructed based on the CNN network, and the loss function is: L ph =||PDEs[on(m,c),m,c]|| p Among them, ||·|| p represents Lp regularization, m represents the model parameters of the geological model, c represents the simulation parameters, sur() represents the production dynamic fitting data obtained according to the geological model and the simulation parameters, and PDEs represents the preset partial differential equation, which is used to express the preset physical constraint relationship satisfied by the production dynamic fitting data output by the preset physical constraint agent network and the geological model and the simulation parameters; The optimization module is used to perform data assimilation using the ES-MDA method based on the production dynamic measured data and the production dynamic fitting data, adjust the input vector, and iteratively update the initial geological model according to the adjusted input vector; when the data error is within a preset error range, the optimal geological model is obtained.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the geological modeling method based on neural network and dynamic data assimilation as claimed in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps of the geological modeling method based on neural network and dynamic data assimilation as described in any one of claims 1 to 6.

Images