Novel intelligent oil and gas reservoir numerical simulation method based on physical mechanism driving

Abstract

The invention discloses a novel intelligent oil and gas reservoir numerical simulation method based on physical mechanism driving, and relates to the cross technical field of computational fluid mechanics and deep learning. Comprising the following steps of oil and gas reservoir mathematical model establishment and grid division, SE attention mechanism feature weight distribution, learnable wavelet operator neural network multi-scale feature extraction, physical constraint scaling of output of a pressure prediction branch and a saturation prediction branch, and performing spatial discretization on the partial differential equations in the S1 by adopting a finite volume method to construct a conservation type discretization format and the like. According to the method, intelligent high-precision numerical simulation of oil and gas reservoir seepage based on a physical mechanism driven neural network is realized, a deep learning method is combined with a physical equation, adaptive extraction of multi-scale features is realized through a learnable wavelet operator, sequential progressive prediction is realized through a transfer learning strategy, and the method is more accurate and efficient. On the premise of ensuring the physical consistency, the spatio-temporal evolution law of the pressure and saturation of the oil and gas reservoir can be efficiently and accurately predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of the intersection of computational fluid dynamics and deep learning, and particularly relates to a new method of intelligent oil and gas reservoir numerical simulation based on physical mechanism driving. BACKGROUND

[0002] In complex physical engineering systems such as petroleum engineering, groundwater resource management, and carbon dioxide geological storage, oil, gas, and water seepage as the most representative multiphase flow phenomenon in porous media, its dynamic behavior can be accurately described by a set of partial differential equations (PDEs). However, the traditional oil and gas reservoir numerical simulation technology faces the problems of low computational efficiency, significant grid dependence, and insufficient processing capacity of high-dimensional uncertainty when solving large-scale spatio-temporal evolution problems, which seriously restricts the real-time optimization and decision-making of oil and gas reservoir development schemes.

[0003] In recent years, the rapid development of artificial intelligence technology has provided a new opportunity to break through the technical bottlenecks of traditional numerical simulation. Raissi et al. proposed a physics-informed neural network (PINNs) that constructs the control equation, initial condition, and boundary condition as a physical loss function to achieve neural network approximation of PDE solutions. However, PINNs still have inherent defects such as limited high-dimensional data processing capacity, slow training convergence speed, and insufficient accuracy and stability of the solution. The subsequently developed neural operator method, especially the wavelet neural operator (WNO), combines the multi-scale decomposition characteristics of wavelet basis functions with the nonlinear mapping ability of neural networks by introducing wavelet multi-resolution analysis theory, and has made significant progress in PDE solution accuracy.

[0004] Although WNO has strong local adaptive ability in dealing with complex physical systems with multi-scale characteristics, local high-frequency components, and significant smoothness (such as oil and gas reservoir seepage), it still has key technical defects such as static unadjustable sub-band information, inability to adaptively enhance or suppress features, and loss of high-frequency information, limiting the ability to capture local high-frequency features. In addition, the pressure equation and saturation equation in the oil, gas, and water three-phase seepage control equation set are strongly coupled, and there are significant differences in physical characteristics and spatial distribution patterns. How to achieve targeted multi-physical field feature extraction in the neural network architecture and effectively embed the physical equation into the network model is still a key scientific problem to be solved.

[0005] The existing technology has the following problems:

[0006] 1. Accuracy and stability of existing neural network solution methods

[0007] The existing physical information neural network has problems of limited high-dimensional data processing capability, slow training convergence speed, insufficient solution precision and stability, and the like when processing complex physical systems such as oil and gas reservoir seepage with strong nonlinearity, multi-scale and significant local high-frequency characteristics, and it is difficult to achieve an effective balance between calculation precision and efficiency.

[0008] 2. The feature extraction capability of the wavelet neural operator (WNO) is insufficient

[0009] The existing wavelet neural operator has the following two key defects although it has multi-scale analysis capability:

[0010] Firstly, the static processing strategy of sub-band information leads to equal weight and isolated learning of different frequency sub-bands, lacks an adaptive regulation mechanism, causes model parameter redundancy and easily introduces noise, and reduces the solving precision and generalization ability; and secondly, only the highest scale wavelet coefficients are parameterized and learned, so that the model is limited in representation and transformation of high-frequency information, cannot effectively capture key high-frequency characteristics such as water injection front and strong pressure gradient near production wells in oil and gas reservoir seepage, and affects the prediction accuracy.

[0011] 3. Difficulty in multi-field coupled solution

[0012] In the mathematical model of oil and gas reservoir seepage, the pressure equation and the saturation equation have strong coupling characteristics, and the pressure field and the saturation field have significant differences in physical characteristics and spatial distribution patterns. How to realize targeted multi-physical field feature extraction in the neural network architecture and effectively embed the physical conservation equation into the network model for coupled solution is a problem faced by the existing technology.

[0013] 4. Efficiency and continuity of time series prediction

[0014] In the time series prediction of oil and gas reservoir development, directly training a long time series neural network easily leads to numerical problems such as gradient disappearance and non-convergence. How to fully utilize the physical continuity of oil and gas reservoir space-time evolution to improve the prediction accuracy while reducing the training time is a problem to be solved by the existing technology. SUMMARY

[0015] The purpose of the present application is to provide a new method of intelligent oil and gas reservoir numerical simulation based on physical mechanism driving, to solve the precision and stability problems of the existing neural network solution method, the feature extraction capability of the wavelet neural operator (WNO) is insufficient, the difficulty in multi-field coupled solution, and the efficiency and continuity of time series prediction.

[0016] To achieve the above purpose, the present application provides the following technical solutions:

[0017] A novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms is characterized by the following steps: establishing a mathematical model and meshing the oil and gas reservoir; assigning feature weights using the SE attention mechanism; extracting multi-scale features from a learnable wavelet operator neural network; applying physical constraints to the outputs of the pressure prediction and saturation prediction branches; constructing a conserved discrete scheme by spatially discretizing the partial differential equations in S1 using the finite volume method; and training the pressure and saturation values ​​for the next time step using a transfer learning temporal training strategy.

[0018] S1. Establishment of Mathematical Model and Mesh Generation for Oil and Gas Reservoirs: A mathematical model for oil and gas reservoirs is established, using the oil-water two-phase flow equations as an example. The formulas are as follows:

[0019]

[0020] in, Represents the oil phase. Represents the water phase. and These represent the saturation, pressure, relative permeability, viscosity, formation volume factor, compressibility factor, and source / sink term of a certain phase, respectively. The relative permeability... It's saturation. nonlinear functions, and These represent rock porosity, rock compressibility, absolute permeability, and capillary force, respectively.

[0021] The oil and gas reservoir model is divided into an N×N Cartesian grid to construct an input feature sample set. The input features include key feature parameters such as absolute permeability, bottom hole pressure at time t, or injection rate.

[0022] S2, SE attention mechanism feature weight allocation: The input features are adaptively weighted using the SE (Squeeze-and-Excitation) attention mechanism. This mechanism includes three operations: Squeeze (global feature compression), Excitation (channel weight learning), and Scale (feature recalibration), where the compression ratio r is a preset value.

[0023] The pressure prediction branch and the saturation prediction branch are each configured with an independent SE module, so that the two types of physical quantities have different attention modes, and targeted multi-physics feature extraction is achieved.

[0024] S3. Multi-scale feature extraction using a learnable wavelet operator neural network: The weighted features after the SE module are mapped to a high-dimensional latent space. A learnable wavelet convolutional layer is established, and a J-layer discrete wavelet transform (DWT) is performed on the input features, decomposing them into: the low-frequency approximate components of the J-th layer and the high-frequency detail components in the horizontal, vertical, and diagonal directions of each layer. Furthermore, the subsequent k layers... The high-frequency detail subbands are subjected to learnable Gabor gating, and the Gabor filter function is applied. Defined as the product of a Gaussian envelope and a cosine carrier wave, it includes learnable parameters such as control bandwidth, frequency, direction, and phase. Through backpropagation adaptive optimization, it applies learnable Gabor-gated detail subbands in the m-th layer l-direction. Represented as:

[0025]

[0026]

[0027] in, For gating strength parameters;

[0028] The learnable multi-scale wavelet expansion with Gabor gating is as follows:

[0029] ,

[0030] Furthermore, a linear mapping transformation is performed on the low-frequency and high-frequency subbands of the last layer, followed by an inverse wavelet transform (IDWT) to reconstruct the processed wavelet coefficients in the spatial domain, forming a multi-scale fused feature map. The wavelet reconstructed features are residually connected with the point convolution features and output through a nonlinear activation function. The learnable wavelet operator layer of the j-th layer is as follows:

[0031] ,

[0032] By stacking the above learnable wavelet operator layers multiple times, multi-scale feature extraction and nonlinear mapping are achieved through iterative deepening. After iterating through multiple layers of learnable wavelet operators, the high-dimensional latent space features are reduced to the original physical space dimension.

[0033] S4. Apply physical constraint scaling to the outputs of the pressure prediction branch and the saturation prediction branch: The final output of saturation is mapped to the sigmoid activation function. Range, pressure output mapped to range ,in, To restrict water saturation, Residual oil saturation, and Set according to the actual reservoir pressure range.

[0034] In the implementation of this invention, the above-mentioned physical constraints ensure that the pressure and saturation values ​​output by the network always fall within the physically feasible range, thus avoiding unreasonable non-physical prediction results.

[0035] S5. The partial differential equations in S1 are spatially discretized using the finite volume method to construct a conserved discretization scheme: the discrete residual form is as follows:

[0036]

[0037] in, Represents the volume of a unit cell; It is the time step; This represents the cross-sectional area between mesh elements i and j. It is the distance between the centers of two units, absolute permeability. By calculating using harmonic averaging, the smoothed L1 norm is defined as the physical loss function, and the total loss function is:

[0038]

[0039] In this invention, the numerical oscillation and accuracy loss are effectively mitigated by using the finite volume method to discretize the control equations, thereby enhancing the model's generalization ability in strongly nonlinear regions.

[0040] S6. Employ a transfer learning-based temporal training strategy to train the pressure and saturation values ​​for the next time step: For a total of T time steps, train T independent learnable wavelet operator neural network models sequentially. Each model is responsible for predicting one time step. For the th time step... Training at each time step: If Initialize network parameters to random values; if Loading the first The model weights trained at the nth time step are used as initialization to implement transfer learning. The physical loss function defined in S5 is calculated, and backpropagation is performed to update the network weights. When training reaches the set maximum number of iterations or the loss is less than a given threshold, network training stops, and the nth time step is output. The pressure and saturation prediction results at each time step are used to iterate through all T time steps to complete the training and prediction.

[0041] The beneficial effects of this invention are as follows: This application realizes high-precision numerical simulation of three-phase seepage in oil and gas reservoirs based on a physical mechanism-driven neural network. This method combines the powerful representational capabilities of deep learning with the conservation constraints of physical equations, achieves adaptive extraction of multi-scale features through learnable wavelet operators, and realizes temporal progressive prediction through transfer learning strategies. It can efficiently and accurately predict the spatiotemporal evolution of pressure and saturation in oil and gas reservoirs while ensuring physical consistency.

[0042] 1. Strong multi-scale feature extraction capability

[0043] This invention, by introducing a learnable mechanism of Gabor gating, achieves adaptive control of different spatial frequency characteristics. It can dynamically enhance or suppress specific frequency components based on the local features of the PDE solution, significantly improving the ability to capture complex spatial structural features. By applying Gabor gating to the high-frequency detail subbands of the last k layers, the ability to represent and transform high-frequency information is further strengthened. This technology has a stronger ability to identify and characterize local high-frequency features in oil and gas reservoir seepage (such as the water injection front and strong pressure gradients near production wells), effectively solving the key defect of high-frequency information loss in existing wavelet neural operators.

[0044] 2. Accurate and efficient multi-field coupling solution

[0045] This invention employs a dual-branch independent SE attention mechanism architecture, configuring independent SE modules and network parameters for the pressure prediction branch and the saturation prediction branch respectively, enabling targeted multi-physics feature extraction. Each branch ensures that the predicted values ​​strictly meet physical feasibility constraints through a physical constraint output mapping mechanism, fully considering the essential differences in physical properties and spatial distribution patterns between the pressure and saturation fields. This design, combining dual-branch independent learning with physical constraints, effectively solves the problem of strongly coupled multi-field input-multi-field output solutions, achieving significant improvements in both solution accuracy and computational efficiency compared to methods predicting multiple physical quantities using a single network architecture.

[0046] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0047] Figure 1 The diagram shows the predicted pressure field and saturation field for day 2 as shown in this invention, along with the error diagram.

[0048] Figure 2 This is a prediction diagram and error diagram of the pressure field and saturation field on day 4 as shown in this invention.

[0049] Figure 3 This is a prediction diagram and error diagram of the pressure field and saturation field on day 6 as shown in this invention.

[0050] Figure 4 This is a prediction diagram and error diagram of the pressure field and saturation field on day 8 as shown in this invention.

[0051] Figure 5 This is a prediction diagram and error diagram of the pressure field and saturation field on day 10 as shown in this invention.

[0052] Figure 6 This is a flowchart illustrating the present invention. Detailed Implementation

[0053] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0055] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0056] This application contains a preferred embodiment.

[0057] 1. Consider a two-dimensional reservoir model with dimensions of 1280m × 1280m, divided into a 64 × 64 uniform grid. The important physical properties of the reservoir are set as follows:

[0058] ,

[0059] The absolute permeability field K is heterogeneously distributed and generated using a Gaussian random field, exhibiting strong spatial variability and effectively simulating the heterogeneous characteristics of real oil reservoirs. Relative permeability is described using the Brooks-Corey model. Five wells are installed within the reservoir: three injection wells with constant flow rate and two production wells with constant pressure. Under initial conditions, the pressure and water saturation fields are uniformly distributed throughout the reservoir. The Peaceman Well Model is used to approximate the pressure gradient near the source / sink, calculated using the following formula:

[0060]

[0061] in, Let be the relative permeability of a certain phase. For absolute penetration rate, For reservoir thickness, Let the viscosity of a certain phase be... For the equivalent radius, Where is the wellbore radius. The skin coefficient is used. Formation parameters and well control information are extracted from the grid to construct the input feature sample set.

[0062] 2. An SE attention mechanism is used to adaptively weight the input features. The compression ratio r is 4, and the pressure branch and saturation branch each use independent SE modules.

[0063] 3. Learnable Wavelet Operator Neural Network for Multi-Scale Feature Extraction. The weighted features after the SE module are mapped to a high-dimensional space, and learnable wavelet convolutional layers are executed. A three-layer Discrete Wavelet Transform (DWT) is performed on the input features, decomposing them into: a low-frequency approximation component in the third layer and high-frequency detail components in the horizontal, vertical, and diagonal directions of each layer. Learnable Gabor gating is applied to the high-frequency detail subbands of the last two layers, while the low-frequency approximation subband remains unchanged. Furthermore, a linear mapping is performed on the low-frequency and high-frequency subbands of the last layer, and an Inverse Wavelet Transform (IDWT) is executed to reconstruct the features. The output is then processed through residual connections and a nonlinear activation function, finally reducing the high-dimensional latent space features back to the original space.

[0064] 4. Apply physical constraints to scale the outputs of the pressure prediction branch and the saturation prediction branch. In this embodiment, The value is 0.2 (bound water saturation). The value is 0.2 (residual oil saturation). 2000psi -1 , 8000psi -1 .

[0065] 5. The finite volume method (FVM) is used to discretize the oil-water two-phase flow equation and construct the physical residual loss function.

[0066] 6. A transfer learning strategy is used for time series training. The total prediction time T is divided into 5 time steps, with each time step Δt being 2 days. The pressure and saturation fields are predicted sequentially for days 2, 4, 6, 8, and 10. The Adam optimizer is used to train the model. Training stops when the number of training iterations reaches the maximum value of 20,000 or the physical loss is less than 0.001, and the pressure and saturation values ​​are output.

[0067] The attached figures show the predicted and error diagrams of the pressure and saturation fields on days 2, 4, 6, 8, and 10.

[0068] In summary, this invention provides a novel numerical simulation method for intelligent oil and gas reservoirs driven by physical mechanisms. This method achieves high-precision numerical simulation of two-phase flow in oil reservoirs based on a neural network driven by physical mechanisms. This method combines the powerful representational capabilities of deep learning with the conservation constraints of physical equations. It achieves adaptive extraction of multi-scale features through learnable wavelet operators and realizes temporal progressive prediction through transfer learning strategies. It can efficiently and accurately predict the spatiotemporal evolution of reservoir pressure and saturation while ensuring physical consistency.

[0069] 1. Strong multi-scale feature extraction capability

[0070] This invention introduces a learnable Gabor gating mechanism to achieve adaptive control of different spatial frequency characteristics. It can dynamically enhance or suppress specific frequency components based on the local characteristics of the PDE solution, significantly improving the ability to capture complex spatial structural features. By applying Gabor gating to the high-frequency detail subbands of the last k layers, the ability to represent and transform high-frequency information is further strengthened. This technology has a stronger ability to identify and characterize local high-frequency features in oil-water two-phase flow (such as the water injection front and strong pressure gradients near production wells), effectively solving the key defect of high-frequency information loss in existing wavelet neural operators.

[0071] 2. Accurate and efficient multi-field coupling solution

[0072] This invention employs a dual-branch independent SE attention mechanism architecture, configuring independent SE modules and network parameters for the pressure prediction branch and the saturation prediction branch respectively, enabling targeted multi-physics feature extraction. Each branch ensures that the predicted values ​​strictly meet physical feasibility constraints through a physical constraint output mapping mechanism, fully considering the essential differences in physical properties and spatial distribution patterns between the pressure and saturation fields. This design, combining dual-branch independent learning with physical constraints, effectively solves the problem of strongly coupled multi-field input-multi-field output solutions, achieving significant improvements in both solution accuracy and computational efficiency compared to methods predicting multiple physical quantities using a single network architecture.

[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms, characterized in that, The process includes the following steps: establishing a mathematical model of the oil and gas reservoir and meshing; assigning feature weights using the SE attention mechanism; extracting multi-scale features from a learnable wavelet operator neural network; applying physical constraints to the outputs of the pressure prediction and saturation prediction branches; using the finite volume method to spatially discretize the partial differential equations in S1 to construct a conserved discrete scheme; and using a transfer learning temporal training strategy to train the pressure and saturation values ​​for the next time step. S1. Establishment of Mathematical Model and Mesh Generation for Oil and Gas Reservoirs: Establish a mathematical model for oil and gas reservoirs and determine the seepage equations. The oil and gas reservoir model is divided into an N×N Cartesian grid to construct an input feature sample set. The input features include key feature parameters such as absolute permeability, bottom hole pressure at time t, or injection rate. S2, SE Attention Mechanism Feature Weight Allocation: The input features are adaptively weighted using the SE attention mechanism, which includes three operations: Squeeze, Excitation and Scale, where the compression ratio r is a preset value; The pressure prediction branch and the saturation prediction branch are each configured with an independent SE module, so that the two types of physical quantities have different attention modes, and targeted multi-physics feature extraction is achieved. S3. Learnable Wavelet Operator Neural Network Multi-Scale Feature Extraction: The weighted features after the SE module are mapped to a high-dimensional latent space. A learnable wavelet convolutional layer is established, and J-layer discrete wavelet transform is performed on the input features, decomposing them into: the low-frequency approximate components of the J-th layer and the high-frequency detail components in the horizontal, vertical, and diagonal directions of each layer. Furthermore, for the last k layers... Learnable Gabor gating is applied to the high-frequency detail subbands, i.e., Gabor-gated learnable multi-scale wavelets. Then, linear mapping transformation is performed on the low-frequency and high-frequency subbands of the last layer. Inverse wavelet transformation is performed to reconstruct the processed wavelet coefficients to the spatial domain, forming a multi-scale fusion feature map. Then, residual connection and nonlinear activation function output are performed. After multiple layers of learnable wavelet operator layers, the high-dimensional latent space features are reduced to the original physical space dimension. S4. Apply physical constraint scaling to the outputs of the pressure prediction branch and the saturation prediction branch: The final output of saturation is mapped to the sigmoid activation function. Range, pressure output mapped to range ,in, To restrict water saturation, Residual oil saturation, and Set according to the actual reservoir pressure range. By using the above physical constraints, we can ensure that the pressure and saturation values ​​output by the network always fall within the physically feasible range, thus avoiding unreasonable non-physical prediction results. S5. The partial differential equations in S1 are spatially discretized using the finite volume method to construct a conserved discretization scheme: Discretizing the control equations using the finite volume method effectively alleviates numerical oscillations and accuracy loss, and enhances the model's generalization ability in strongly nonlinear regions. S6. Employ a transfer learning-based temporal training strategy to train the pressure and saturation values ​​for the next time step: For a total of T time steps, train T independent learnable wavelet operator neural network models sequentially. Each model is responsible for predicting one time step. For the th time step... Training at each time step: If Initialize network parameters to random values; if Loading the first The model weights trained at the nth time step are used as initialization to implement transfer learning. The physical loss function defined in S5 is calculated, and backpropagation is performed to update the network weights. When training reaches the set maximum number of iterations or the loss is less than a given threshold, network training stops, and the nth time step is output. The pressure and saturation prediction results at each time step are used to iterate through all T time steps to complete the training and prediction.

2. The novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms as described in claim 1, characterized in that, The oil-water two-phase flow equations in S1 are as follows: in, Represents the oil phase. Represents the water phase. and These represent the saturation, pressure, relative permeability, viscosity, formation volume factor, compressibility factor, and source / sink term of a certain phase, respectively. The relative permeability... It's saturation. nonlinear functions, and These represent rock porosity, rock compressibility, absolute permeability, and capillary force, respectively.

3. The novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms as described in claim 1, characterized in that, Learnable Gabor gating in S3 involves Gabor filtering functions. Defined as the product of a Gaussian envelope and a cosine carrier wave, it includes learnable parameters such as control bandwidth, frequency, direction, and phase. Through backpropagation adaptive optimization, it applies learnable Gabor-gated detail subbands in the m-th layer l-direction. Represented as: ， ， in, This refers to the gating strength parameter.

4. The novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms as described in claim 1, characterized in that, The learnable multi-scale wavelet expansion of Gabor-gated S3 is as follows: ， Subsequently, a linear mapping transformation is performed on the low-frequency and high-frequency subbands of the last layer. An inverse wavelet transform is then executed to reconstruct the processed wavelet coefficients in the spatial domain, forming a multi-scale fused feature map. The wavelet-reconstructed features are then residually connected with point convolutional features and output through a nonlinear activation function. The learnable wavelet operator layer of the j-th layer is as follows: By stacking the aforementioned learnable wavelet operator layers multiple times, multi-scale feature extraction and nonlinear mapping are achieved through iterative deepening. After multiple layers of wavelet operator iteration, the high-dimensional latent space features are reduced to the original physical space dimension.

5. The novel numerical simulation method for intelligent oil and gas reservoirs based on physical mechanisms as described in claim 1, characterized in that, The residual form of the finite volume discretization in S5 is as follows: in, Represents the volume of a unit cell; It is the time step; This represents the cross-sectional area between mesh elements i and j. It is the distance between the centers of two units, absolute permeability. By calculating using harmonic averaging, the smoothed L1 norm is defined as the physical loss function, and the total loss function is: 。

Images