JIT-based reservoir parameter end-to-end inversion method

Abstract

The invention discloses a reservoir parameter end-to-end inversion method based on JIT, and relates to the technical field of engineering geology. The method comprises the following steps: preprocessing reservoir geological attribute data and injection-production well production dynamic data to construct a data set, then constructing a reservoir parameter inversion model coupled by a pure image Transformer model, a weighted overlapped local block convolution fusion model and a sampling generation module, training the reservoir parameter inversion model by using samples in the data set until the reservoir parameter inversion model is qualified, and finally obtaining the reservoir parameter inversion model after the reservoir parameter inversion model is qualified. And directly inputting the production dynamic data into the reservoir parameter inversion model, generating reservoir geological parameters matched with the production dynamic data, and realizing geological parameter inversion. According to the method, the contradiction between large-scale modeling and detail precision is solved, direct mapping from observation data to a geological parameter field is realized, and a reliable technical means is provided for development of underground reservoirs.

Description

Technical Field

[0001] This invention relates to the field of engineering geology, specifically to a JIT-based end-to-end reservoir parameter inversion method. Background Technology

[0002] Geological parameters, as core indicators characterizing the structure and physical properties of underground reservoirs, with key parameters such as porosity and permeability directly determining the reservoir's fluid storage capacity and seepage patterns, significantly impact the design rationality and operational safety of underground engineering projects such as oil and gas reservoir development and geothermal resource utilization. Therefore, accurate inversion of geological parameters is a crucial foundation for improving the efficiency of underground reservoir development, optimizing engineering schemes, and reducing development risks.

[0003] Traditional geological parameter inversion methods primarily rely on deterministic optimization algorithms or stochastic sampling techniques, such as gradient descent adjoint methods and Markov chain Monte Carlo methods. These methods require iterative solutions to complex forward models, with each simulation involving large-scale numerical calculations of partial differential equations, resulting in extremely high computational costs and making them unsuitable for the efficient inversion requirements of large-scale 3D geological models in practical engineering. Furthermore, traditional geological parameter inversion methods are highly dependent on the initial model and are prone to getting trapped in local optima in scenarios with strong nonlinearity and high parameter dimensionality, making uncertainty assessment of the inversion results difficult and limiting their reliable application under complex geological conditions.

[0004] To alleviate the computational efficiency bottleneck of traditional geological parameter inversion methods and improve inversion efficiency, surrogate model methods are gradually being applied to the field of reservoir geological parameter inversion. The core of this method is to replace complex reservoir forward modeling models with simplified models, reducing the frequency of large-scale numerical calculations, thereby lowering inversion costs and improving efficiency. However, this method still requires calling optimization algorithms for solving, and the inversion process remains cumbersome. It fails to fundamentally solve the problem of the complex process of traditional geological parameter inversion methods and cannot meet the high-precision and high-efficiency inversion requirements in complex reservoir scenarios.

[0005] In recent years, with the development of artificial intelligence technology, generative inversion methods such as those based on generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models have been gradually applied to this field. These methods, by learning the mapping relationship between prior geological structures and observational data, can quickly generate geological models that conform to statistical characteristics, providing a new path to solve the efficiency bottleneck of traditional methods. However, the generative methods used in existing technologies still have significant limitations. For example, GAN training is unstable and prone to pattern collapse; VAEs generate samples of limited quality; and while diffusion models produce good results, their sampling efficiency is low. Furthermore, when dealing with large-scale, high-resolution 3D geological models, there are problems such as high training difficulty, a large number of parameters, and the need for massive amounts of high-quality samples. Instability and convergence are prone to occur in high-dimensional parameter spaces. Even with the introduction of dimensionality reduction or complex pre-training strategies, it often comes at the cost of losing geological details and reducing model flexibility, making it difficult to efficiently adapt to actual inversion tasks at different work areas and scales.

[0006] Therefore, it is urgent to propose an end-to-end reservoir parameter inversion method based on JIT, which uses a pure image transformer method to invert geological parameters, improves the detail accuracy and spatial continuity of the geological parameter inversion results, and realizes a fine characterization of the underground reservoir structure and physical properties. Summary of the Invention

[0007] This invention aims to solve the above-mentioned problems and proposes a JIT-based end-to-end reservoir parameter inversion method. It integrates the pure image transformer method with weighted overlapping local block convolution to achieve accurate inversion of geological parameters, forming a brand-new end-to-end geological parameter inversion method. This method resolves the contradiction between large-scale modeling and detailed accuracy, ensuring both the efficiency and accuracy of geological parameter inversion. It significantly improves the detail realism and spatial continuity of the inversion results, providing a reliable technical means for geological modeling, scheme optimization, and risk management in underground reservoir development.

[0008] The present invention adopts the following technical solution: The JIT-based end-to-end reservoir parameter inversion method includes the following steps: Step 1: Based on reservoir geological attribute data and injection-production well production dynamic data, construct a production dynamic data sample set and a reservoir geological attribute sample set. After preprocessing the production dynamic data sample set and the reservoir geological attribute sample set, construct the dataset. Step 2: Construct a pure image Transformer model to process reservoir geological attribute samples in the dataset; Step 3: Construct a weighted overlapping local block convolutional fusion model to fuse and reconstruct fusion features and reconstructed image to obtain the reconstruction parameter field; Step 4: Construct a sampling generation module to generate geological parameters; Step 5: Couple the pure image Transformer model constructed in Step 2, the weighted overlapping local block convolutional fusion model constructed in Step 3, and the sampling generation module constructed in Step 4 to obtain the reservoir parameter inversion model. Train the reservoir parameter inversion model using the dataset, input the production dynamic data into the reservoir parameter inversion model, and use the reservoir parameter inversion model to generate a posterior reservoir geological attribute sample set to achieve geological parameter inversion.

[0009] Preferably, step 1 includes the following sub-steps: Step 1.1: Obtain reservoir geological parameters and geological attribute data through seismic exploration, well logging interpretation, and core analysis; based on the reservoir geological attribute data, construct multiple prior geological models using stochastic geological modeling methods, and then perform numerical simulations using each prior geological model. The injection process was simulated to obtain the production dynamic data of injection and production wells corresponding to each prior geological model. A production dynamic data sample set was constructed based on the reservoir geological attribute data of all prior geological models. , , For the first The production dynamics data of injection and production wells from each prior geological model are used; then, a reservoir geological attribute sample set is constructed based on the reservoir geological attribute data of all prior geological models. , , For the first Reservoir geological attribute data from a prior geological model; Step 1.2: Normalize the production dynamic data of injection and production wells in the production dynamic data sample set and the reservoir geological attribute data in the reservoir geological attribute sample set, respectively, to obtain the normalized production dynamic data sample set. and reservoir geological property sample set , construct the dataset.

[0010] Preferably, step 2 includes the following sub-steps: Step 2.1: Use linear interpolation to normalize the reservoir geological attribute sample set. Noise-adding processing is performed to obtain the noisy reservoir geological attribute data. , ,in, For the set of real numbers, For batch size, , , These are reservoir geological attribute data. The number of floors, height, and width; Step 2.2, processing the noisy reservoir geological attribute data. Preprocessing is performed, and reservoir geological attribute data are processed using an overlapping sliding window. Divided into A local block; for the segmented reservoir geological attribute data Convolutional feature extraction and feature serialization are performed sequentially to obtain serialized local block features. for: ; in, ; In the formula, These are the local block features after serialization; For flattening operation; For local block feature maps; It is a two-dimensional convolution function; Step 2.3: Generate corresponding positional codes based on each local block, and then apply these codes to the serialized local block features. After adding positional encoding, the serialized local block features with added positional encoding are then processed. Perform rotational position encoding to obtain the rotated local block features. for: ; in, ; In the formula, , All are spatial location indices of local blocks; For located Local block location encoding; For rotational position encoding function; Step 2.4: The normalized injection-production well production dynamic data... and time step As a conditional embedding, it guides the generation of inversion parameters. A Transformer encoder, including a multi-head self-attention module and a feedforward neural network, is constructed to learn the mapping between injection-production well production dynamics data and geological parameter fields. The Transformer encoder is then used to obtain sequence features. ; Step 2.5, the sequence features after passing through the Transformer encoder Reconstructing the reservoir geological attribute data yields: ; In the formula, For the reconstructed reservoir geological attribute data; This is a reshaping function.

[0011] Preferably, step 2.4 includes the following sub-steps: Step 2.4.1, according to the time step Generate temporal embedding vectors ,get: ; In the formula, It is a multilayer perceptron; Based on the production dynamics data of injection and production wells Generate observation data embedding vectors ,get: ; Temporal embedding vector and observation data embedding vector Adding them together yields the comprehensive condition vector. , ; Step 2.4.2, based on the comprehensive condition vector The normalized parameters of the Transformer encoder are generated as follows: ; In the formula, For the adaptive scaling factor of the multi-head self-attention module; For the adaptive offset of the multi-head self-attention module; To control the gating weights of self-attention residual connections; This is the adaptive scaling factor for the feedforward network module; This is the adaptive offset of the feedforward network module; To control the gating weights of the residual connections in the feedforward network; Construct a multi-head self-attention module to process position-encoded local block features. After performing the projection transformation, the parts are split and the attention is calculated, resulting in: ; in, ; ; ; In the formula, For attention; , , These are the projected query vector, key vector, and value vector, respectively. It is the transpose matrix; For each attention head, the dimension is: , , These are the learnable projection matrices corresponding to the query vector, key vector, and value vector, respectively. It is a normalized exponential function; Step 2.4.3: For the location-encoded local block features... By implementing residual connectivity and gating mechanisms, and dynamically adjusting the scale and offset of local block features based on injection-production well production dynamic data, condition-aware feature normalization is achieved, resulting in: ; In the formula, This refers to the local block features after condition-aware normalization. It is the root mean square normalization function; For element-wise multiplication; Local block features after conditional awareness normalization The input is fed into a multi-head self-attention module and compared with the local block features initially input into the multi-head self-attention module. Gated residual connections are performed, and then the local block features obtained through the attention mechanism and gated residual connections are input into the feedforward neural network for adaptive feature selection processing to obtain sequence features. ; The sequence features Represented as: ; in, ; In the formula, This refers to local block features connected through attention mechanisms and gated residuals; Linear layer; This is the activation function.

[0012] Preferably, step 3 includes the following sub-steps: Step 3.1: Establish a learnable weight matrix for local block fusion reconstruction, resulting in: ; In the formula, As a learnable weight matrix, through The function will learn the weight matrix Learnable parameters in Value constraints to ; By traversing the positions of all local blocks and calculating the spatial mapping position of each local block, the weights of overlapping local blocks are accumulated and then normalized to reconstruct the parameter field, resulting in: ; in, ; ; In the formula, The reconstructed parameter field after normalization; Accumulated images based on local blocks; It is a function for maximizing the value; This is the weighted cumulative matrix; It is the numerical stability constant; The total number of reservoir geological attribute data after reshaping; For the first in the complete image The matrix corresponding to each local block; It is the transpose matrix; For the reshaping of the first Individual reservoir geological attribute data values; Step 3.2: Multi-scale feature extraction is performed on the normalized reconstructed parameter field using a multi-layer convolutional network to obtain the reconstructed fusion features. A weighted overlapping local block convolutional fusion model is then constructed. The reconstructed fusion features are combined with the reconstructed image using the weighted overlapping local block convolutional fusion model to obtain the weighted reconstructed convolutional fusion parameter field. The expression for the weighted overlapping local block convolutional fusion model is: ; In the formula, To reconstruct the parameter field after weighted convolution fusion; These are the residual weighting coefficients; To reconstruct the fusion features.

[0013] Preferably, step 4 includes the following sub-steps: Step 4.1: Construct a sampling generation module and use it to calculate the weighted reconstructed convolutional fusion parameter field. Flow vector field The calculation formula is: ; In the formula, This is the reservoir geological attribute data after noise addition; For time steps; Step 4.2: Based on the Euler integral method, invert reservoir geological attribute data to generate posterior reservoir geological attribute samples and construct a posterior reservoir geological attribute sample set. The formula for inverting and calculating the reservoir geological attribute data is as follows: ; In the formula, This serves as a sample of the geological properties of the reservoir for future research. It is randomly sampled Gaussian noise.

[0014] Preferably, in step 5, a pure image Transformer model, a weighted overlapping local block convolutional fusion model, and a sampling generation module are sequentially encapsulated to establish a reservoir parameter inversion model. The reservoir parameter inversion model is trained using matching production dynamic data samples and reservoir geological attribute samples in the dataset until it meets the preset training qualification conditions. Then, the trained reservoir parameter inversion model is used to perform geological parameter inversion. The production dynamic data collected in the field is input into the reservoir parameter inversion model. After the pure image Transformer model in the reservoir parameter inversion model predicts and reconstructs the reservoir geological attribute data, the parameter field is reconstructed through the weighted overlapping local block convolutional fusion model. Then, the sampling generation module is used to sample the weighted reconstructed convolutional fusion parameter field to obtain posterior reservoir geological attribute samples, generating a posterior reservoir geological attribute sample set, and inverting the geological parameters corresponding to the production dynamic data collected in the field.

[0015] The present invention has the following beneficial effects: (1) This invention proposes an end-to-end reservoir parameter inversion method based on JIT. For the first time, it innovatively combines pure image Transformer with weighted overlapping local block convolution fusion technology to construct a brand-new end-to-end reservoir geological parameter inversion model. It effectively solves the problem that large-scale modeling and detailed accuracy are difficult to balance in traditional reservoir geological parameter inversion methods, and realizes efficient and high-precision inversion of large-scale three-dimensional geological parameter fields.

[0016] (2) The present invention proposes a JIT-based end-to-end inversion method for reservoir parameters. This method addresses the problem of high dimensionality and training difficulty of large-scale three-dimensional geological parameter fields by adopting an adaptive local block strategy for geological structures. By intelligently identifying the spatial correlation features of geological structures, the complex parameter field is divided into a series of local blocks with spatial contextual correlation. This not only significantly reduces the input dimension and training complexity of the neural network, but also fully preserves the global correlation of the geological structure, laying the foundation for efficient modeling.

[0017] Meanwhile, the method of this invention innovatively adopts a weighted overlapping local block convolution fusion mechanism in the reconstruction stage: by adaptively adjusting the contribution of different local blocks through a learnable weight matrix, the features of key geological areas are enhanced in a targeted manner; at the same time, the convolution operation is used to capture the spatial dependency between local blocks, smooth the boundary effect generated by the splicing between blocks, effectively make up for the loss of details caused by the processing of local blocks, and significantly improve the matching degree between the geological parameter inversion results and the actual geological conditions.

[0018] (3) The present invention proposes an end-to-end reservoir parameter inversion method based on JIT. This method adopts an end-to-end framework without complex manual intervention and multi-stage iteration, realizing direct mapping from observation data to geological parameter field. It achieves the optimal balance between computational feasibility and detail authenticity, providing strong technical support for geological fine modeling, real-time optimization of development schemes and long-term safety risk management involved in underground reservoir development. It is conducive to promoting the transformation of geological parameter inversion from experience-driven to data-intelligent driven paradigm, and has important engineering application value and broad industry promotion prospects. Attached Figure Description

[0019] Figure 1 This is a flowchart of the JIT-based end-to-end reservoir parameter inversion method of the present invention.

[0020] Figure 2 This is a schematic diagram of the permeability field along the x-direction in a three-dimensional saline aquifer carbon sequestration model; inj1, inj2, inj3, and inj4 are all injection wells in the diagram.

[0021] Figure 3 The diagram shows the permeability field in the x-direction of a real reservoir model. In the diagram, (a) is the permeability field in the x-direction of the third stratigraphic layer in the real reservoir model, (b) is the permeability field in the x-direction of the fourth stratigraphic layer in the real reservoir model, (c) is the permeability field in the x-direction of the fifth stratigraphic layer in the real reservoir model, and (d) is the permeability field in the x-direction of the sixth stratigraphic layer in the real reservoir model.

[0022] Figure 4 The figures show the inversion results of the permeability field of the three-dimensional saline aquifer carbon sequestration model. In the figure, (a) is the inversion result of the permeability field of the third stratigraphic structure in the x-direction of the three-dimensional saline aquifer carbon sequestration model, (b) is the inversion result of the permeability field of the fourth stratigraphic structure in the x-direction of the three-dimensional saline aquifer carbon sequestration model, (c) is the inversion result of the permeability field of the fifth stratigraphic structure in the x-direction of the three-dimensional saline aquifer carbon sequestration model, and (d) is the inversion result of the permeability field of the sixth stratigraphic structure in the x-direction of the three-dimensional saline aquifer carbon sequestration model.

[0023] Figure 5 The diagram shows the pressure fitting comparison of the injection wells in this embodiment. In the diagram, (a) is a schematic diagram of the comparison results of injection well I1, (b) is a schematic diagram of the comparison results of injection well I2, (c) is a schematic diagram of the comparison results of injection well I3, and (d) is a schematic diagram of the comparison results of injection well I4.

[0024] Figure 6The following is a comparison diagram of the pressure fitting of the monitoring wells in this embodiment; in the figure, (a) is a schematic diagram of the comparison results of monitoring well M1, (b) is a schematic diagram of the comparison results of monitoring well M2, (c) is a schematic diagram of the comparison results of monitoring well M5, and (d) is a schematic diagram of the comparison results of monitoring well M6.

[0025] Figure 7 The figures below are comparison diagrams of CO2 saturation fitting for monitoring wells in this embodiment. In the figures, (a) is a schematic diagram of the comparison results of CO2 saturation fitting for monitoring well M1, (b) is a schematic diagram of the comparison results of CO2 saturation fitting for monitoring well M2, (c) is a schematic diagram of the comparison results of CO2 saturation fitting for monitoring well M4, and (d) is a schematic diagram of the comparison results of CO2 saturation fitting for monitoring well M7.

[0026] Figure 8 The figure shows the physical field obtained by numerical simulation of the real reservoir model at 1500 days. In the figure, (a) is the gas phase saturation field obtained by numerical simulation of the real reservoir model at 1500 days, and (b) is the molar density field of CO2 obtained by numerical simulation of the real reservoir model at 1500 days.

[0027] Figure 9 The figure shows the physical field of the three-dimensional saline aquifer carbon sequestration model obtained by inversion using the method of the present invention and simulated under the same working conditions for 1500 days. In the figure, (a) is the gas phase saturation field of the model obtained by inversion using the method of the present invention and simulated by numerical simulation for 1500 days, and (b) is the molar density field of CO2 obtained by inversion using the model obtained by inversion using the method of the present invention and simulated by numerical simulation for 1500 days.

[0028] Figures 5-7 In the diagram, the gray line represents the prior production dynamic data of injection-production wells (referred to as prior samples), the red line represents the actual production dynamic data of injection-production wells (referred to as actual observation data), and the blue line represents the simulated production dynamic data of injection-production wells based on the inversion results (referred to as posterior samples). Detailed Implementation

[0029] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0030] Example 1 This invention proposes an end-to-end inversion method for reservoir parameters based on Just Image Transformer (JIT), such as... Figure 1 As shown, the specific steps include: Step 1: Based on reservoir geological attribute data and injection-production well production dynamic data, construct a production dynamic data sample set and a reservoir geological attribute sample set. After preprocessing the production dynamic data sample set and the reservoir geological attribute sample set, construct the dataset, including the following sub-steps: Step 1.1: Obtain reservoir geological parameters and geological attribute data through seismic exploration, well logging interpretation, and core analysis. This data will provide a realistic and high-precision constraint basis for the subsequent construction of the reservoir parameter inversion model, ensuring that the reservoir parameter inversion model can accurately reflect the geological characteristics and spatial distribution patterns of the actual reservoir.

[0031] Based on reservoir geological attribute data, multiple prior geological models were constructed using a stochastic geological modeling method to reflect the spatial variability and uncertainty of reservoir parameters; then, each prior geological model was used in a numerical simulator for... The injection process was simulated to obtain the production dynamic data of injection and production wells corresponding to each prior geological model.

[0032] A production dynamics data sample set was constructed based on the reservoir geological attribute data of all prior geological models. , , For the first The production dynamics data of injection and production wells from each prior geological model are used; then, a reservoir geological attribute sample set is constructed based on the reservoir geological attribute data of all prior geological models. , , For the first Reservoir geological attribute data from a priori geological model.

[0033] Step 1.2: Normalize the production dynamic data of injection and production wells in the production dynamic data sample set and the reservoir geological attribute data in the reservoir geological attribute sample set, respectively, to obtain the normalized production dynamic data sample set. and reservoir geological property sample set , construct the dataset.

[0034] Step 2: Construct a pure image Transformer model to process reservoir geological attribute samples in the dataset.

[0035] The pure image Transformer model, as a lightweight diffusion generation method, is based on the visual Transformer architecture. It directly predicts clean data rather than noise or high noise content through a large-size image patch processing mode. It does not rely on pre-training, additional loss functions, or word segmenters, and achieves efficient sampling by relying on stream matching methods and solutions.

[0036] The pure image Transformer model construction process includes the following sub-steps: Step 2.1: Use linear interpolation to normalize the reservoir geological attribute sample set. Noise-adding processing is performed to obtain the noisy reservoir geological attribute data. , ,in, For the set of real numbers, For batch size, , , These are reservoir geological attribute data. The number of layers, height, and width.

[0037] Specifically, the formula for calculating the noise reduction of the reservoir geological attribute data is as follows: ; In the formula, This is the reservoir geological attribute data after noise addition; For time steps; It is Gaussian noise. Gaussian distribution function middle This is the noise scale, with a default value of 1.0.

[0038] Step 2.2, processing the noisy reservoir geological attribute data. Preprocessing is performed, and reservoir geological attribute data are processed using an overlapping sliding window. It is divided into multiple local blocks, and the total number of local blocks is... , ,in, , In the formula, , These represent the number of local blocks in the vertical and horizontal directions, respectively. , These are the height and width of the local block, respectively. , These are the sliding step sizes in the vertical and horizontal directions, respectively.

[0039] First, local spatial features within each local block are extracted, and the pixel-level information of the original parameters is transformed into a high-dimensional feature representation. This unifies the feature dimension of each local block and adapts it to the input requirements of a pure image Transformer model. Then, the two-dimensional local block feature map is transformed into a one-dimensional sequence to adapt to the one-dimensional sequence input format that the pure image Transformer model can process, while preserving the integrity of the local block features. In this embodiment, this is achieved by processing the segmented reservoir geological attribute data... Convolutional feature extraction and feature serialization are performed sequentially to obtain serialized local block features. for: ; in, ; In the formula, These are the local block features after serialization; For flattening operation; For local block feature maps; It is a two-dimensional convolution function.

[0040] Step 2.3: To imbue the serialized local block features with spatial location information, enabling the pure image Transformer model to recognize the relative positions of local blocks and avoid losing spatial topological relationships due to serialization, it is necessary to... Add position encoding.

[0041] In this embodiment, the local block features are first serialized. Add position encoding Then, the serialized local block features with added position encoding... Perform rotational position encoding to enhance the ability to perceive the relative position of local blocks and obtain the features of the rotated local blocks. for: ; in, ; In the formula, , All are spatial location indices of local blocks; For located Local block location encoding; This is a rotation position encoding function.

[0042] Step 2.4: The normalized injection-production well production dynamic data... and time step As a conditional embedding, it guides the generation of inversion parameters and constructs a Transformer encoder including a multi-head self-attention module and a feedforward neural network to learn the mapping between injection-production well production dynamics data and geological parameter fields. The specific embedding method is as follows: Step 2.4.1, according to the time step Generate temporal embedding vectors ,get: ; In the formula, It is a multilayer perceptron.

[0043] Based on the production dynamics data of injection and production wells Generate an observation data embedding vector, the observation data embedding vector for: ; In the formula, Embed vectors into the observation data.

[0044] Temporal embedding vector and observation data embedding vector The summation yields the comprehensive condition vector, calculated using the following formula: ; In the formula, This is the comprehensive condition vector.

[0045] Step 2.4.2, based on the comprehensive condition vector The normalized parameters of the Transformer encoder are generated as follows: ; In the formula, For the adaptive scaling factor of the multi-head self-attention module; For the adaptive offset of the multi-head self-attention module; To control the gating weights of self-attention residual connections; This is the adaptive scaling factor for the feedforward network module; This is the adaptive offset of the feedforward network module; To control the gating weights of the residual connections in the feedforward network.

[0046] Construct a multi-head self-attention module to process position-encoded local block features. After performing the projection transformation, the parts are split and the attention is calculated, resulting in: ; in, ; ; ; In the formula, For attention; , , These are the projected query vector, key vector, and value vector, respectively. It is the transpose matrix; For each attention head, the dimension is: , , These are the learnable projection matrices corresponding to the query vector, key vector, and value vector, respectively. It is a normalized exponential function.

[0047] Step 2.4.3: For the location-encoded local block features... By implementing residual connectivity and gating mechanisms, and dynamically adjusting the scale and offset of local block features based on injection-production well production dynamic data, condition-aware feature normalization is achieved, resulting in: ; In the formula, This refers to the local block features after condition-aware normalization. It is the root mean square normalization function; This is for element-wise multiplication.

[0048] Local block features after conditional awareness normalization The input is fed into a multi-head self-attention module and compared with the local block features initially input into the multi-head self-attention module. Gated residual connections are performed, and then the local block features obtained through the attention mechanism and gated residual connections are input into the feedforward neural network for adaptive feature selection processing to obtain sequence features. .

[0049] The sequence features The calculation formula is: ; in, ; In the formula, This refers to local block features connected through attention mechanisms and gated residuals; Linear layer; This is the activation function.

[0050] Step 2.5, the sequence features after passing through the Transformer encoder Reconstructing the reservoir geological attribute data yields: ; In the formula, For the reconstructed reservoir geological attribute data; This is a reshaping function.

[0051] Step 3: Construct a weighted overlapping local block convolutional fusion model to fuse and reconstruct features with the reconstructed image to obtain the reconstruction parameter field. This weighted overlapping local block convolutional fusion model is the core module connecting the output of the Transformer in the pure image Transformer model with the reconstruction of the complete geological parameter field. It is used to solve the boundary effects and detail loss problems in the local block segmentation and reconstruction process, achieving high-precision fusion of local block features into the global parameter field.

[0052] The construction process of the weighted overlapping local block convolutional fusion model includes the following sub-steps: Step 3.1: Establish a learnable weight matrix for local block fusion reconstruction, resulting in: ; In the formula, As a learnable weight matrix, through The function will learn the weight matrix Learnable parameters in Value constraints to .

[0053] In this embodiment, the fusion weights of local blocks at different locations are adaptively adjusted by training a learnable weight matrix, while the weights of the learnable parameters are constrained in the [0,1] interval to ensure the physical meaning of the learnable parameter weights.

[0054] Traverse the positions of all local blocks and calculate the spatial mapping position of each local block. The formula for calculating the spatial mapping position of a local block is as follows: ; ; In the formula, , All are spatial location indices of local blocks; , , , These correspond to the column start position, column end position, row start position, and row end position of the local block, respectively. , These are the height and width of the local block, respectively. , These are the sliding step sizes in the vertical and horizontal directions, respectively.

[0055] After summing the weights of the overlapping local blocks and normalizing them, the parameter field is reconstructed to obtain: ; in, ; ; In the formula, The reconstructed parameter field after normalization; Accumulated images based on local blocks; It is a function for maximizing the value; This is the weighted cumulative matrix; Here is the numerical stability constant, which takes the value of ; The total number of reservoir geological attribute data after reshaping; For the first in the complete image The matrix corresponding to each local block; It is the transpose matrix; For the reshaping of the first Individual reservoir geological attribute data values.

[0056] Step 3.2: Use a multi-layer convolutional network to extract multi-scale features from the reconstructed parameter field after normalization, and obtain the reconstructed fusion features. A weighted overlapping local block convolutional fusion model is constructed, and the reconstructed fusion features are combined with the reconstructed image using the weighted overlapping local block convolutional fusion model to obtain the parameter field after weighted reconstruction convolutional fusion. .

[0057] The expression for the weighted overlapping local block convolutional fusion model is: ; In the formula, To reconstruct the parameter field after weighted convolution fusion; These are the residual weighting coefficients; To reconstruct the fusion features.

[0058] In summary, this embodiment first traverses all local block locations and calculates the spatial mapping position of each local block to ensure that the features of each local block are accumulated to the correct position, avoiding topological chaos in the parameter field caused by spatial mapping errors. Then, the features of each local block are weighted according to learnable weights and accumulated to the corresponding spatial position, while recording the accumulated weight values ​​to provide a basis for the subsequent normalization processing of the reconstructed parameter field. After preserving the feature information of overlapping regions, the accumulated result is normalized to eliminate the numerical inflation and compression problems caused by the weight accumulation in overlapping regions, mapping the accumulated reconstructed parameter field to a reasonable numerical range and ensuring the physical rationality of the reconstruction result. Finally, a multi-layer convolutional network is used to process the normalized reconstructed parameter field. Multi-scale feature extraction is performed to obtain reconstructed and fused features. This improves the detail integrity of the reconstructed parameter field, and combines the reconstructed fusion features with the reconstructed image through residual connections to obtain the weighted reconstructed convolutional fusion parameter field. The weighted overlapping local block convolution fusion model was constructed.

[0059] Step 4: Construct a sampling generation module to generate geological parameters, which includes the following sub-steps: Step 4.1: Construct a sampling generation module and use it to calculate the weighted reconstructed convolutional fusion parameter field. Flow vector field The calculation formula is: ; In the formula, This is the reservoir geological attribute data after noise addition; For time steps.

[0060] Step 4.2: Based on the Euler integral method, invert the reservoir geological attribute data to generate posterior reservoir geological attribute samples, resulting in: ; In the formula, This serves as a sample of the geological properties of the reservoir for future research. It is randomly sampled Gaussian noise.

[0061] Based on the generated posterior reservoir geological attribute samples, a posterior reservoir geological attribute sample set is constructed.

[0062] Step 5: Couple the pure image Transformer model constructed in Step 2, the weighted overlapping local block convolutional fusion model constructed in Step 3, and the sampling generation module constructed in Step 4 to obtain the reservoir parameter inversion model. Train the reservoir parameter inversion model using the dataset until the reservoir parameter inversion model reaches the preset training qualification conditions. Then, input the production dynamic data into the reservoir parameter inversion model and use the reservoir parameter inversion model to generate a posterior reservoir geological attribute sample set to achieve geological parameter inversion.

[0063] Specifically, a pure image Transformer model, a weighted overlapping local block convolutional fusion model, and a sampling generation module are sequentially encapsulated to construct an integrated reservoir parameter inversion model. The reservoir parameter inversion model is used to invert geological parameters. The production dynamic data collected in the field is input into the pure image Transformer model. After the pure image Transformer model predicts and reconstructs the reservoir geological attribute data, the parameter field is reconstructed through the weighted overlapping local block convolutional fusion model. Then, the sampling generation module is used to sample the weighted reconstructed convolutional fusion parameter field to obtain posterior reservoir geological attribute samples, generate a posterior reservoir geological attribute sample set, and invert the geological parameters.

[0064] Example 2 To verify the feasibility and superiority of the JIT-based end-to-end reservoir parameter inversion method described in Example 1, this example applies the JIT-based end-to-end reservoir parameter inversion method to a three-dimensional saline aquifer carbon sequestration model. The three-dimensional saline aquifer carbon sequestration model includes 4 injection wells and 5 monitoring wells. This model consists of seven formation layers: the first layer is a saline aquifer, the second layer is a caprock, and the third to seventh layers are reservoirs. It contains a total of 573,440 grid blocks of 128×128×5m each, representing a 20m×20m×5m formation. The uncertain parameter is the permeability field of the three-dimensional saline aquifer carbon sequestration model, while the observation data includes pressure and saturation measurements at the monitoring locations. The production history spans 1500 days, with a total of 50 time steps, and 21 fitting indices are used. Figure 2 This is a schematic diagram of the permeability field along the x-direction in a three-dimensional saline aquifer carbon sequestration model. Figure 3The x-direction permeability field represents the third, fourth, fifth, and sixth stratigraphic layers in a real reservoir model.

[0065] The geological parameters of the three-dimensional saline aquifer carbon sequestration model were inverted using the JIT-based end-to-end reservoir parameter inversion method described in Example 1.

[0066] First, a dataset was constructed, which included a carbon sequestration geological parameter field and corresponding injection-production well production dynamic data. Following the process in step 1, 1200 samples were constructed. After completing multi-source data acquisition, prior model construction, and numerical simulation, the data was normalized to obtain 1200 sets of samples paired with reservoir geological attribute data and production dynamic data.

[0067] Following the process in step 2, a pure image Transformer model is constructed to process reservoir geological attribute samples in the dataset. The normalized geological parameters are noise-added, and the model is segmented into local blocks through overlapping sliding windows and convolutional features are extracted. Two-dimensional sine and cosine and rotation position encodings are added, and the comprehensive conditional embedding of production dynamic data and time steps is fused. The Transformer encoder block is constructed to learn the mapping relationship between dynamic data and geological parameter fields, and the sequence features are reshaped into local block format.

[0068] Following the process in step 3, a weighted overlapping local block convolutional fusion architecture is constructed, generating a learnable weight matrix and constraining it to the range [0,1]. The spatial mapping positions of the local blocks are calculated through traversal, and the weight accumulation and normalization of the overlapping local blocks are completed. Fusion features are extracted through a multi-scale convolutional network, and the complete reconstructed geological parameter field is obtained by combining residual connections. Furthermore, a sampling generation module is constructed to generate geological parameters.

[0069] Finally, the pure image Transformer model constructed in step 2, the weighted overlapping local block convolutional fusion model constructed in step 3, and the sampling generation module constructed in step 4 are coupled together to obtain the reservoir parameter inversion model. The reservoir parameter inversion model is trained using the dataset constructed in step 1, and the geological parameters are inverted and solved.

[0070] By preprocessing the actual carbon sequestration injection-production well dynamic data and inputting it into a pure image Transformer model, and combining it with a weighted overlapping local block convolution fusion model to complete feature reconstruction, followed by inverse normalization processing, the final output is a three-dimensional carbon sequestration geological parameter field at a specified time step. Then, the sampling generation module in step 4 samples this data to generate an inverted permeability field, thus obtaining the permeability field of the three-dimensional saline aquifer carbon sequestration model, as shown below. Figure 4 As shown, comparison Figure 3 It can be seen that the permeability field obtained by inverting production dynamic data using the method of the present invention is highly consistent with the real field in terms of spatial distribution pattern and numerical range.

[0071] To verify the historical fitting effect of the method of this invention, a carbon sequestration numerical simulator was first used to numerically simulate the reservoir permeability field predicted by the method of this invention, generating dynamic production data (i.e., pressure and saturation data) of injection and production wells corresponding to the prediction model. Then, the simulated data was compared with the actual dynamic production data of injection and production wells in the reservoir. The comparison results of fitting from selected injection wells and monitoring wells are presented, such as... Figures 5-7 As shown, Figure 5 This is a pressure fitting comparison chart of the injection well in this embodiment, used to show the degree of agreement between the actual reservoir injection well pressure data and the pressure data simulated by the inversion model; Figure 6 This is a pressure fitting comparison chart of the monitoring well in this embodiment, used to show the matching between the actual reservoir monitoring well pressure and the simulated pressure obtained by the inversion model; Figure 7 This is a comparison chart of the CO2 saturation fitting of the monitoring wells in this embodiment, used to intuitively reflect the consistency between the actual reservoir monitoring wells and the CO2 saturation of the inverted model.

[0072] Depend on Figures 5-7 It is evident that the oil and water well production dynamic curves obtained using the method of this invention closely match the changing trends and numerical fluctuations of the actual reservoir production dynamic curves, demonstrating excellent fitting accuracy. Furthermore, comparing the physical fields obtained through numerical simulations of the actual model and the inverted model in the embodiments of this invention, such as... Figure 8 and Figure 9 As shown. Figure 8 The physical field at 1500 days is obtained from numerical simulation of the real model. Figure 9 The inverted model obtained in this embodiment of the invention is used to numerically simulate the physical field at 1500 days. (Comparison) Figure 8 and Figure 9 It can be seen that the inversion results of the method of the present invention are highly consistent with the actual field in terms of both numerical values ​​and morphology. This shows that the method of the present invention has significant effectiveness in inverting reservoir geological parameters under large-scale three-dimensional geological conditions.

[0073] In summary, the method of this invention can effectively improve the detail accuracy and spatial continuity of geological parameter inversion results, achieve a fine characterization of underground reservoir structure and physical properties, provide technical support for underground reservoir development, and has broad practical engineering application value.

[0074] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A JIT-based reservoir parameter end-to-end inversion method, characterized in that, Includes the following steps: Step 1: Based on reservoir geological attribute data and injection-production well production dynamic data, construct a production dynamic data sample set and a reservoir geological attribute sample set. After preprocessing the production dynamic data sample set and the reservoir geological attribute sample set, construct the dataset. Step 2: Construct a pure image Transformer model to process reservoir geological attribute samples in the dataset; Step 3: Construct a weighted overlapping local block convolutional fusion model to fuse and reconstruct fusion features and reconstructed image to obtain the reconstruction parameter field; Step 4: Construct a sampling generation module to generate geological parameters; Step 5: Couple the pure image Transformer model constructed in Step 2, the weighted overlapping local block convolutional fusion model constructed in Step 3, and the sampling generation module constructed in Step 4 to obtain the reservoir parameter inversion model. Train the reservoir parameter inversion model using the dataset, input the production dynamic data into the reservoir parameter inversion model, and use the reservoir parameter inversion model to generate a posterior reservoir geological attribute sample set to achieve geological parameter inversion.

2. The JIT-based end-to-end reservoir parameter inversion method according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1.1: Obtain reservoir geological parameters and geological attribute data through seismic exploration, well logging interpretation, and core analysis; based on the reservoir geological attribute data, construct multiple prior geological models using stochastic geological modeling methods, and then perform numerical simulations using each prior geological model. The injection process was simulated to obtain the production dynamic data of injection and production wells corresponding to each prior geological model. A production dynamic data sample set was constructed based on the reservoir geological attribute data of all prior geological models. , , For the first The production dynamics data of injection and production wells from each prior geological model are used; then, a reservoir geological attribute sample set is constructed based on the reservoir geological attribute data of all prior geological models. , , For the first Reservoir geological attribute data from a prior geological model; Step 1.2: Normalize the production dynamic data of injection and production wells in the production dynamic data sample set and the reservoir geological attribute data in the reservoir geological attribute sample set, respectively, to obtain the normalized production dynamic data sample set. and reservoir geological property sample set , construct the dataset.

3. The JIT-based end-to-end reservoir parameter inversion method according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1: Use linear interpolation to normalize the reservoir geological attribute sample set. Noise-adding processing is performed to obtain the noisy reservoir geological attribute data. , ,in, For the set of real numbers, For batch size, , , These are reservoir geological attribute data. The number of floors, height, and width; Step 2.2, processing the noisy reservoir geological attribute data. Preprocessing is performed, and reservoir geological attribute data are processed using an overlapping sliding window. Divided into A local block; for the segmented reservoir geological attribute data Convolutional feature extraction and feature serialization are performed sequentially to obtain serialized local block features. for: ； in, ； In the formula, These are the local block features after serialization; For flattening operation; For local block feature maps; It is a two-dimensional convolution function; Step 2.3: Generate corresponding positional codes based on each local block, and then apply these codes to the serialized local block features. After adding positional encoding, the serialized local block features with added positional encoding are then processed. Perform rotational position encoding to obtain the rotated local block features. for: ； in, ； In the formula, , All are spatial location indices of local blocks; For located Local block location encoding; For rotational position encoding function; Step 2.4: The normalized injection-production well production dynamic data... and time step As a conditional embedding, it guides the generation of inversion parameters. A Transformer encoder, including a multi-head self-attention module and a feedforward neural network, is constructed to learn the mapping between injection-production well production dynamics data and geological parameter fields. The Transformer encoder is then used to obtain sequence features. ; Step 2.5, the sequence features after passing through the Transformer encoder Reconstructing the reservoir geological attribute data yields: ； In the formula, For the reconstructed reservoir geological attribute data; This is a reshaping function.

4. The JIT-based end-to-end reservoir parameter inversion method according to claim 3, characterized in that, Step 2.4 includes the following sub-steps: Step 2.4.1, according to the time step Generate temporal embedding vectors ,get: ； In the formula, It is a multilayer perceptron; Based on the production dynamics data of injection and production wells Generate observation data embedding vectors ,get: ； Temporal embedding vector and observation data embedding vector Adding them together yields the comprehensive condition vector. , ; Step 2.4.2, based on the comprehensive condition vector The normalized parameters of the Transformer encoder are generated as follows: ； In the formula, For the adaptive scaling factor of the multi-head self-attention module; For the adaptive offset of the multi-head self-attention module; To control the gating weights of self-attention residual connections; This is the adaptive scaling factor for the feedforward network module; This is the adaptive offset of the feedforward network module; To control the gating weights of the residual connections in the feedforward network; Construct a multi-head self-attention module to process position-encoded local block features. After performing the projection transformation, the parts are split and the attention is calculated, resulting in: ； in, ； ； ； In the formula, For attention; , , These are the projected query vector, key vector, and value vector, respectively. It is the transpose matrix; For each attention head, the dimension is... , , These are the learnable projection matrices corresponding to the query vector, key vector, and value vector, respectively. It is a normalized exponential function; Step 2.4.3: For the location-encoded local block features... By implementing residual connectivity and gating mechanisms, and dynamically adjusting the scale and offset of local block features based on injection-production well production dynamic data, condition-aware feature normalization is achieved, resulting in: ； In the formula, This refers to the local block features after condition-aware normalization. It is the root mean square normalization function; For element-wise multiplication; Local block features after conditional awareness normalization The input is fed into a multi-head self-attention module and compared with the local block features initially input into the multi-head self-attention module. Gated residual connections are performed, and then the local block features obtained through the attention mechanism and gated residual connections are input into the feedforward neural network for adaptive feature selection processing to obtain sequence features. ; The sequence features Represented as: ； in, ； In the formula, This refers to local block features connected through attention mechanisms and gated residuals; Linear layer; This is the activation function.

5. The JIT-based end-to-end reservoir parameter inversion method according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Establish a learnable weight matrix for local block fusion reconstruction, resulting in: ； In the formula, As a learnable weight matrix, through The function will learn the weight matrix Learnable parameters in Value constraints to ; By traversing the positions of all local blocks and calculating the spatial mapping position of each local block, the weights of overlapping local blocks are accumulated and then normalized to reconstruct the parameter field, resulting in: ； in, ； ； In the formula, The reconstructed parameter field after normalization; Accumulated images based on local blocks; It is a function for maximizing the value; This is the weighted cumulative matrix; It is the numerical stability constant; The total number of reservoir geological attribute data after reshaping; For the first in the complete image The matrix corresponding to each local block; It is the transpose matrix; For the reshaping of the first Individual reservoir geological attribute data values; Step 3.2: Multi-scale feature extraction is performed on the normalized reconstructed parameter field using a multi-layer convolutional network to obtain the reconstructed fusion features. A weighted overlapping local block convolutional fusion model is then constructed. The reconstructed fusion features are combined with the reconstructed image using the weighted overlapping local block convolutional fusion model to obtain the weighted reconstructed convolutional fusion parameter field. The expression for the weighted overlapping local block convolutional fusion model is: ； In the formula, To reconstruct the parameter field after weighted convolution fusion; These are the residual weighting coefficients; To reconstruct the fusion features.

6. The JIT-based end-to-end reservoir parameter inversion method according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Construct a sampling generation module and use it to calculate the weighted reconstructed convolutional fusion parameter field. Flow vector field The calculation formula is: ； In the formula, This is the reservoir geological attribute data after noise addition; For time steps; Step 4.2: Based on the Euler integral method, invert reservoir geological attribute data to generate posterior reservoir geological attribute samples and construct a posterior reservoir geological attribute sample set. The formula for inverting and calculating the reservoir geological attribute data is as follows: ； In the formula, This serves as a sample of the geological properties of the reservoir for future research. It is randomly sampled Gaussian noise.

7. The JIT-based end-to-end reservoir parameter inversion method according to claim 1, characterized in that, In step 5, the pure image Transformer model, the weighted overlapping local block convolutional fusion model, and the sampling generation module are sequentially encapsulated to establish a reservoir parameter inversion model. The reservoir parameter inversion model is trained using matching production dynamic data samples and reservoir geological attribute samples in the dataset until it meets the preset training qualification conditions. Then, the geological parameters are inverted using the trained reservoir parameter inversion model. The production dynamic data collected in the field is input into the reservoir parameter inversion model. After the pure image Transformer model in the reservoir parameter inversion model predicts and reconstructs the reservoir geological attribute data, the parameter field is reconstructed through the weighted overlapping local block convolutional fusion model. Then, the sampling generation module samples the weighted reconstructed convolutional fusion parameter field to obtain posterior reservoir geological attribute samples, generating a posterior reservoir geological attribute sample set. The geological parameters corresponding to the production dynamic data collected in the field are then inverted.

Images