Transform-CNN (Convolutional Neural Network) architecture-based large-scale reservoir model reconstruction method and equipment

Abstract

The invention provides a large-scale reservoir model reconstruction method and device based on a Transform-CNN architecture, and relates to the field of three-dimensional geological modeling and deep learning, and the method comprises the steps: S1, constructing an initial MAE network based on a Transform network and a convolutional layer; s2, training the initial MAE network based on progressive generation and partition simulation to obtain a large-scale MAE network; and S3, inputting the reservoir model to be reconstructed into the large-scale MAE network to obtain a reconstructed reservoir model. According to the method, the initial MAE network is constructed based on the Transform network and the convolutional layer, so that the parameter quantity of the neural network is effectively reduced, and the modeling scale is improved. An initial MAE network is trained in a progressive generation and partition simulation mode, a large-scale MAE network is finally obtained by iteratively training different scales of MAE networks, the large-scale MAE network can consider the size of a complete model, the overall space structure of a large-scale reservoir model is effectively restored, and the problem of error propagation in a traditional progressive generation method is solved.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional geological modeling and deep learning, and in particular to a large-scale reservoir model reconstruction method and device based on a Transformer-CNN architecture. Background Art

[0002] In recent years, many researchers have combined deep learning with reservoir modeling, giving rise to a series of reservoir modeling methods based on deep generative models. Among them, the convolutional neural network (CNN) extracts multi-scale features of training images layer by layer through multi-layer convolution transformation and pooling, which can well characterize the geometric morphology and spatial structure characteristics of the reservoir. The generative adversarial network (GAN) achieves realistic and diverse image generation through adversarial generation between the generator and the discriminator. The combination of convolutional neural network and adversarial generation method (referred to as CNNs-GANs architecture) provides a more effective solution for the fine characterization of reservoir heterogeneous structure.

[0003] In the modeling process based on deep learning methods, the designed neural network model adaptively learns the feature representation in the training data during the training process of the model, and establishes a data mapping path from input to output. This mapping is reused when new data is input to achieve efficient modeling. The reservoir modeling method based on deep learning has a more efficient modeling process and more refined characterization effect than the traditional geological modeling method, but it also has certain bottlenecks in some aspects. For example, CNN relies on the large receptive field formed by the deep stack of convolution operations to extract the long-range dependencies of the reservoir model. However, long-distance dependencies are easy to disappear or difficult to learn in the process of multiple convolution stacking, which makes the extraction of the global feature space of the model still a bottleneck. Secondly, the robustness of conditional simulation of many existing deep learning methods is insufficient. Specifically, it is manifested in: insufficient conditional fitting ability. The indirect conditional generation method used by GAN guides the change of feature distribution of the generated domain by taking conditional data as additional constraint information. The generated model is difficult to be completely faithful to the conditional data; conditional training is unstable. The commonly used adversarial generation training method often has problems such as mode collapse and gradient disappearance; the conditional generalization performance is insufficient. The direct conditional simulation method, by directly finding the mapping relationship between specific conditional data and the model feature distribution, may cause the generated samples to overfit on the original conditional data, resulting in performance degradation when new conditional data is input.

[0004] In reservoir exploration engineering, improving the scale of reservoir modeling is crucial for finely characterizing reservoir structure and properties and supporting deep-earth resource spatial exploration. Existing deep learning reservoir modeling methods can be roughly divided into two types from the perspective of model generation: holistic generation and progressive generation. The holistic generation deep learning method takes the complete model as the generation target and directly establishes an end-to-end mapping between the original input and the complete model. However, since the neural network learns the feature representation of the original model through complex network structure and parameters, it is often necessary to make a trade-off between modeling effect and modeling scale. The reservoir modeling method based on MAE is a holistic generation strategy. Although it can generate a more detailed reservoir model, its complex parameterization limits the improvement of modeling scale. The progressive generation deep learning method first generates local model blocks of the complete model, and then combines them into a complete model through splicing and other means. However, this method is difficult to fully consider the global characteristics of the complete reservoir model, so it is not as good as the holistic generation method in terms of modeling effect. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a large-scale reservoir model reconstruction method and equipment based on the Transformer-CNN architecture, which is used to solve the problem that the existing modeling process based on deep learning methods has many bottlenecks, and the existing overall generation and progressive generation methods have unsatisfactory model reconstruction effects.

[0006] The present invention provides a large-scale reservoir model reconstruction method based on Transformer-CNN architecture, comprising:

[0007] S1: Build the initial MAE network based on the Transformer network and convolutional layers;

[0008] S2: Train the initial MAE network based on progressive generation and partition simulation to obtain a large-scale MAE network;

[0009] S3: Input the reservoir model to be reconstructed into the large-scale MAE network to obtain the reconstructed reservoir model.

[0010] Preferred:

[0011] The initial MAE network includes an encoder and a decoder;

[0012] The encoder consists of N Transformer networks connected in sequence;

[0013] The decoder consists of K downsampling convolutional layers and K upsampling convolutional layers connected sequentially.

[0014] Preferred:

[0015] The Transformer network includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence;

[0016] The input of the Transformer network is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.

[0017] Preferred:

[0018] Both the downsampling convolution layer and the upsampling convolution layer are 2D convolution layers.

[0019] Preferably, step S2 specifically comprises:

[0020] S21: obtaining a reference reservoir model TI and a reservoir model to be reconstructed Z, the size of the reservoir model to be reconstructed Z is H×L×W, and initializing the size of the conditional data to H×1×1;

[0021] S22: partition the reservoir model Z to be reconstructed to obtain multiple regions, region Z i The dimensions are H'×L'×W';

[0022] S23: The region set of conditional data in the reservoir model Z to be reconstructed is Z ie ;

[0023] S24: Divide the reference reservoir model TI into a first training set T1 of size H×L×W and a second training set T2 of size H'×L'×W';

[0024] S25: Train the initial MAE network through the first training set T1 to obtain a network model M1 with a size of H'×L'×W', a mask rate of m1, and calculate the loss function I1;

[0025] S26: Train the initial MAE network using the second training set T2 to obtain a network model M2 with a size of H'×1×1, a mask rate of m2, and calculate the loss function I2;

[0026] S27: Reconstruct the region set Z through conditional data and network model M2 ie The area where the number of pixels reaches the mask rate m1 required by the network model M1, and the reconstructed area set Z is obtained ie’ ;

[0027] S28: According to the position of the conditional data in the reservoir model Z to be reconstructed, the reconstruction area set Z ie’ Restore to the position where it should be;

[0028] S29: Reconstruct the region set Z ie’ As new condition data, the region set Z will be reconstructedie’ Input the network model M1 to obtain the reconstructed reservoir model;

[0029] S210: Repeat steps S22-S29 until the loss function I1 and the loss function I2 are both less than the preset values, and use the final network model M1 as the large-scale MAE network.

[0030] Preferred:

[0031] The formula for loss function I1 is calculated as:

[0032]

[0033] Among them, λ1 is the first weight parameter, λ2 is the second weight parameter, A1 is the first activation matrix, It is the loss between the reconstructed reservoir model output by the network model M1 and the labeled reservoir model.

[0034] Preferred:

[0035] The formula for loss function I2 is calculated as:

[0036]

[0037] Among them, λ3 is the third weight parameter, λ4 is the fourth weight parameter, A2 is the second activation matrix, It is the loss between the reconstructed reservoir model output by the network model M2 and the labeled reservoir model.

[0038] A large-scale reservoir model reconstruction device based on Transformer-CNN architecture comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the large-scale reservoir model reconstruction method based on Transformer-CNN architecture.

[0039] The present invention has the following beneficial effects:

[0040] (1) The present invention constructs an initial MAE network based on the Transformer network and convolutional layers, effectively reduces the length of the attention sequence by using the convolutional layer (CNN) as a decoder, and introduces some prior knowledge. This improvement not only reduces the computational cost of the network, but also introduces local and translational isovariant inductive biases, thereby improving the modeling accuracy and effect. In the overall generation scheme, the number of neural network parameters is effectively reduced, the modeling scale is improved, and better results are shown than the pure Transformer network in small sample data sets.

[0041] (2) The present invention trains the initial MAE network through progressive generation and partition simulation, and finally obtains a large-scale MAE network by iteratively training MAE networks of different scales. The large-scale MAE network can take into account the size of the complete model and effectively restore the overall spatial structure of the large-scale reservoir model; and the large-scale MAE network can use the model generated by the small-scale MAE network as conditional data, which can effectively reduce the length of the attention sequence and overcome the error propagation problem in the traditional progressive generation method.

[0042] (3) The present invention can be applied in various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software, and has strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;

[0044] Figure 2 It is the structural diagram of the initial MAE network;

[0045] Figure 3 It is a comparison between the traditional training process based on pattern splicing and the training process based on progressive generation and partition simulation of the present invention;

[0046] Figure 4 Three sets of data sets were divided for testing and evaluating the method proposed in the present invention;

[0047] Figure 5 It is a comparison of the experimental cases and statistical characteristics of the three-dimensional binary attribute lithofacies model;

[0048] Figure 6 For the three-dimensional discrete attribute lithofacies model experiment and its statistical characteristics comparison;

[0049] Figure 7 To compare the partition simulation experiments and their statistical characteristics;

[0050] Figure 8 This is a structural diagram of the device according to an embodiment of the present invention;

[0051] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0053] Reference Figure 1The present invention provides a large-scale reservoir model reconstruction method based on the Transformer-CNN architecture, which can achieve accurate characterization of reservoir heterogeneous structures under sparse conditional data, and at the same time break through the limitations of the neural network model in the scale and size of three-dimensional modeling.

[0054] include:

[0055] S1: Build the initial MAE network based on the Transformer network and convolutional layers;

[0056] Furthermore, the initial MAE network includes an encoder and a decoder;

[0057] The encoder consists of N Transformer networks connected in sequence;

[0058] The decoder consists of K downsampling convolutional layers and K upsampling convolutional layers connected sequentially.

[0059] Specifically, the structure of the initial MAE network is as follows Figure 2 As shown in the figure, the Transformer-CNN (convolutional layer) hybrid architecture is used to build the initial MAE network. The decoder of the traditional MAE network is changed from the transformer network to the CNN architecture. The convolutional layer is used to optimize the traditional decoder for the following two reasons:

[0060] 1) In the encoder and decoder of the asymmetric structure, the decoder receives the input of the original size and restores the output of the original size. The encoder only needs to introduce part of the input after the random mask and output the latent vector of the partial size. Since the algorithm complexity of self-attention modeling is quadratic, the encoder has effectively reduced the sequence length through random masking, so using convolutional layers instead of decoders can minimize the algorithm complexity.

[0061] 2) The Transformer network has been proven to outperform CNN in many large-scale trainings by integrating many excellent designs, such as self-attention and residual structure. However, compared with CNN, the Transformer lacks local inductive bias, so its generalization ability is worse than CNN when modeling with small samples. Through CNN, we try to incorporate some inductive bias to make the model more generalized and have stronger short-term modeling capabilities, thereby ensuring the quality of model generation.

[0062] Furthermore, batch normalization (BN) is used in the initial MAE network to unify data distribution and improve network stability; dropout is used in each layer to prevent overfitting; the activation function uses the linear rectifier unit Relu function with sparse activation to effectively avoid the gradient vanishing problem.

[0063] Furthermore, the Transformer network includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence;

[0064] The input of the Transformer network is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.

[0065] Furthermore, both the downsampling convolution layer and the upsampling convolution layer are 2D convolution layers.

[0066] Specifically, the decoder uses a fully convolutional network structure similar to U-Net but without jump connections. It does not use pooling layers, but implements downsampling and upsampling processing through convolution and transposed convolution. The advantage of this is that it can adapt to models of different scales without fixing the dimension of the convolution layer, and gradually learns the spatial features of the model from local to overall in the process of downsampling to upsampling. Experiments have shown that both 2D convolution and 3D volume convolution are effective in reconstructing reservoir models, so this method chooses to use 2D convolution instead of 3D volume convolution, thereby unifying the network structure of 3D reconstruction and 2D reconstruction.

[0067] S2: Train the initial MAE network based on progressive generation and partition simulation to obtain a large-scale MAE network;

[0068] Furthermore, the initial MAE network based on the Transformer-CNN hybrid architecture optimizes and improves the maximum scale of a single end-to-end mapping from the perspective of overall model generation. However, overall generation is still limited by network parameters when facing large-scale and large-scale modeling needs. Improving network performance can only increase the scale of modeling to a certain extent, and progressive generation is required to achieve a method that is not limited by network parameters. However, existing methods based on pattern splicing are still lacking in the ability to restore the global scale of the model.

[0069] Therefore, the present invention trains the initial MAE network based on progressive generation and partition simulation, and replaces the process of pattern splicing by training a larger-scale MAE network, so that the progressive generation method can take into account the large-scale spatial structure of the complete model. The larger-scale MAE network can use the model generated by the small-scale MAE network as conditional data, which can effectively reduce the length of the attention sequence. The method based on partition simulation realizes the global features that characterize the complete model through network training, and can also obtain a certain anti-interference ability through network training, thereby minimizing the impact of the generation error of the small-scale model on the full-scale model.

[0070] Step S2 is specifically as follows:

[0071] S21: obtaining a reference reservoir model TI and a reservoir model to be reconstructed Z, the size of the reservoir model to be reconstructed Z is H×L×W, and initializing the size of the conditional data to H×1×1;

[0072] S22: partition the reservoir model Z to be reconstructed to obtain multiple regions, region Z i The dimensions are H'×L'×W';

[0073] S23: The region set of conditional data in the reservoir model Z to be reconstructed is Z ie ;

[0074] S24: Divide the reference reservoir model TI into a first training set T1 of size H×L×W and a second training set T2 of size H'×L'×W';

[0075] S25: Train the initial MAE network through the first training set T1 to obtain a network model M1 with a size of H'×L'×W', a mask rate of m1, and calculate the loss function I1;

[0076] Furthermore, the formula for loss function I1 is calculated as:

[0077]

[0078] Among them, λ1 is the first weight parameter, λ2 is the second weight parameter, A1 is the first activation matrix, It is the loss between the reconstructed reservoir model output by the network model M1 and the labeled reservoir model.

[0079] S26: Train the initial MAE network using the second training set T2 to obtain a network model M2 with a size of H'×1×1, a mask rate of m2, and calculate the loss function I2;

[0080] Furthermore, the formula for the loss function I2 is calculated as:

[0081]

[0082] Among them, λ3 is the third weight parameter, λ4 is the fourth weight parameter, A2 is the second activation matrix, It is the loss between the reconstructed reservoir model output by the network model M2 and the labeled reservoir model.

[0083] S27: Reconstruct the region set Z through conditional data and network model M2 ie The area where the number of pixels reaches the mask rate m1 required by the network model M1, and the reconstructed area set Z is obtained ie’ ;

[0084] S28: According to the position of the conditional data in the reservoir model Z to be reconstructed, the reconstruction area set Z ie’ Restore to the position where it should be;

[0085] S29: Reconstruct the region set Z ie′ As new condition data, the region set Z will be reconstructed ie′ Input the network model M1 to obtain the reconstructed reservoir model;

[0086] S210: Repeat steps S22-S29 until the loss function I1 and the loss function I2 are both less than the preset values, and use the final network model M1 as the large-scale MAE network.

[0087] Specifically, steps S22 to S26 are based on the Transformer-CNN hybrid architecture to optimize and improve the maximum scale of a single end-to-end mapping from the perspective of overall model generation; and step S27 is a progressive generation optimization method based on a partition simulation strategy.

[0088] like Figure 3 The figure shows a comparison between the progressive generation schemes based on pattern splicing and partition simulation in the present invention. Both progressive generation schemes require training a small-scale MAE network. After that, the traditional method based on pattern splicing obtains a complete large-scale model by splicing the results generated by the small-scale network along the simulation path. The method based on partition simulation realizes the global features that characterize the complete model through network training, and can also obtain a certain anti-interference ability through network training, thereby minimizing the impact of the generation error of the small-scale model on the full-scale model.

[0089] S3: Input the reservoir model to be reconstructed into the large-scale MAE network to obtain the reconstructed reservoir model.

[0090] Further, such as Figure 4 As shown, three sets of data sets are divided for testing and evaluating the method proposed in the present invention: Figure 4 (a) is a 180×150×120 scale three-dimensional binary attribute lithofacies model and its 96×96×96 scale data set after division. Figure 4 (b) is a 180×150×120 scale three-dimensional continuous attribute lithofacies model and its 96×96×96 scale data set after division. Figure 4 (c) is a 128×128×64 dataset and a 16×16×16 dataset after the three-dimensional binary attribute lithofacies model is divided. The three datasets are used to perform three sets of experiments, the first two of which are to test the effect of the Transformer-CNN hybrid architecture optimization network, and the latter is to test the effect of the partition simulation method.

[0091] like Figure 5 Shown is a three-dimensional binary attribute lithofacies model experimental case and its statistical characteristics comparison: Figure 5 (a) shows the three reference models of the 96×96×96 scale three-dimensional binary attribute lithofacies model and the corresponding reconstruction results. It can be observed that the reconstructed model of the proposed method can well match the heterogeneous structure of the original reference model in terms of visual appearance, and the connectivity in the details (white box) is better than that of the original reference model. Figure 5 (b) is the conditional simulation result display of a 96×96×96 scale three-dimensional binary attribute lithofacies model at a mask rate of 90% (left: reference model, middle: conditional data, right: reconstruction result). The reconstructed model also conforms well to the constraints of the conditional data. These visual appearance performances preliminarily prove that the overall generation method proposed in the present invention can improve the modeling scale without affecting the modeling quality. Figure 5 (c) is the variogram of one reference model and the corresponding 20 reconstruction results of the three-dimensional binary lithofacies model. The variograms of the reference model and the reconstructed model almost overlap, which proves from a statistical point of view that the reconstruction effect of this method on this data set is very good. Figure 5 (d) is the connectivity curve of 20 reference models and the corresponding 20 reconstruction results of the 3D binary lithofacies model. The connectivity curves of the reference model and its 20 reconstruction results in the X and Y directions are almost overlapped, further confirming the reconstruction performance of this method. Figure 5 (e) is the MDS diagram of 20 reference models of the three-dimensional binary lithofacies model and the corresponding 20 reconstruction results. It can be found that the blue points and the red points are distributed in the range of -500 to 500, and the distances between many blue points and red points are very close, which indicates that the intrinsic characteristics of the reference model and the reconstruction results are very similar. Figure 5 (f) is a box plot of 20 reference models of the three-dimensional binary lithofacies model and the corresponding 20 reconstruction results. Although the maximum, minimum and median values ​​of the box plots are somewhat different, their distances are basically around 0.005, which is mainly due to the close distribution of the 20 randomly selected reference models. Therefore, the box plot can prove that the reconstruction results are very similar to the original models in terms of attribute proportions.

[0092] like Figure 6 Shown is a comparison of the three-dimensional discrete attribute lithofacies model experiment and its statistical characteristics: Figure 6 (a) shows three reference models of the 96×96×96 scale three-dimensional continuous attribute lithofacies model and the corresponding reconstruction results. The reconstruction effect shows rich detail features, the boundary between the yellow attribute and the orange attribute is clear, and the line shape between the orange attribute and the blue attribute is obvious and flows smoothly. This shows that the overall generation method in the present invention has a very good reconstruction effect. Figure 6(b) is the conditional simulation result of a 96×96×96 scale three-dimensional continuous attribute lithofacies model at a masking rate of 90% (left: reference model, middle: conditional data, right: reconstruction result). The reconstructed model and the conditional data can match well, proving the reliability of the conditional simulation of this method. Figure 6 (c) is the variation function curve of a reference model of the three-dimensional continuous attribute lithofacies model and 20 implementation results of the CNN decoder and the Transformer decoder. Compared with the red curve, the blue curve is closer to the black curve in terms of line trend and distance, which proves that the present invention has a strong learning ability for the heterogeneous characteristics of the reservoir model. Figure 6 (d) is the MDS diagram of 20 reference models of the three-dimensional continuous attribute lithofacies model and the corresponding 20 realization results. The red and blue points are evenly distributed in the area of ​​-200 to 200, proving that the generated reconstruction results are highly reliable. Figure 6 (e) is a probability density histogram of 20 reference models of the three-dimensional continuous attribute lithofacies model and the corresponding 20 reconstruction results falling in 50 attribute intervals, showing that the attribute ratios of the reference model and the reconstruction results are very similar in both value range and maximum value, proving that the CNN decoder can indeed incorporate inductive bias and can surpass the Transformer decoder in small sample learning.

[0093] like Figure 7 Shown is the partition simulation experiment and its statistical characteristics comparison. Figure 7 (a) is a complete 128×128×64 reference model and original conditional data. The original conditional data is divided into several regions of 64×16×16 scale, and the region containing the conditional data is selected and further divided into small image blocks of 16×16×16. Figure 7 (b) shows the 16×16×16 conditional data and reconstruction results after partial partitioning. The reconstructed results can restore the basic structural features, but the local connectivity is not good enough. This may be because the scale of the network reconstruction model is too small, so the slight differences in the details can greatly affect the overall reconstruction effect. Therefore, all the small pattern blocks reconstructed by the partitions are subsequently reshaped back to the original space as conditional data, and the large-scale model is used for the final reconstruction. Figure 7 (c) The partition reconstruction result is filled back into the conditional data of 128×128×64 pixels and the complete reconstruction result. Although the effect in the partition conditional reconstruction area is not ideal, the final reconstruction result is very similar to the reference model in appearance, proving that the connectivity features of the reference model have been learned through the network. If the pattern splicing method is used, then due to error propagation, the subtle gap between the small pattern block and the original model will gradually expand to a larger global gap. This proves that this method has the ability to take into account the global feature representation effect of the model and correct local errors. Figure 7 (d) is the variogram curve of the 16 selected cross sections of the reference model and the 16 corresponding cross sections of the reconstruction results. The blue curve and the red curve have basically similar trends and similar upper and lower ranges, which proves that the internal structure of the generated results of the partition simulation is basically consistent with the original model in terms of spatial variability. Figure 7 (e) is the connectivity curve of the 16 selected cross sections of the reference model and the 16 corresponding cross sections of the reconstruction results, which can also prove that the internal structure of the generated results of the partition simulation is basically consistent with the original model in terms of spatial variability. Figure 7 (f) is the MDS graph of the 16 selected cross sections of the reference model and the 16 corresponding cross sections of the reconstruction results. The blue points and the red points are basically close to each other, and it can be considered that the generated cross sections can match each other. Figure 7 (g) is a box plot of the 16 selected reference model sections and the 16 corresponding reconstruction results. The maximum, median and minimum values ​​are slightly different, about 0.2-0.4. The reason for this is probably that the small model blocks of the partition reconstruction are not connected enough, resulting in slightly different attribute proportions. In general, statistical analysis shows that the results generated by the partition simulation method can meet the spatial heterogeneity of the original model and can effectively reduce the uncertainty of reservoir simulation.

[0094] The above three experimental cases show that the present invention effectively reduces the attention sequence length by replacing the Transformer decoder with CNN and introduces some prior knowledge, which can not only improve the modeling scale but also improve the modeling quality; the present invention replaces pattern splicing with network training, takes into account the global characteristics of the original model, and can effectively prevent the propagation of modeling errors.

[0095] See also Figure 8 , Figure 8 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a large-scale reservoir model reconstruction device 401 based on the Transformer-CNN architecture, a processor 402 and a storage medium 403.

[0096] A large-scale reservoir model reconstruction device 401 based on the Transformer-CNN architecture: The large-scale reservoir model reconstruction device 401 based on the Transformer-CNN architecture implements the large-scale reservoir model reconstruction method based on the Transformer-CNN architecture.

[0097] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the large-scale reservoir model reconstruction method based on the Transformer-CNN architecture.

[0098] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the large-scale reservoir model reconstruction method based on the Transformer-CNN architecture.

[0099] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0100] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0101] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A large-scale reservoir model reconstruction method based on Transformer-CNN architecture, characterized in that: include: S1: Build the initial MAE network based on the Transformer network and convolutional layers; S2: Train the initial MAE network based on progressive generation and partition simulation to obtain a large-scale MAE network; S3: Input the reservoir model to be reconstructed into the large-scale MAE network to obtain the reconstructed reservoir model.

2. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 1 is characterized in that: The initial MAE network includes an encoder and a decoder; The encoder consists of N Transformer networks connected in sequence; The decoder consists of K downsampling convolutional layers and K upsampling convolutional layers connected sequentially.

3. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 2 is characterized in that: The Transformer network includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence; The input of the Transformer network is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.

4. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 2 is characterized in that: Both the downsampling convolution layer and the upsampling convolution layer are 2D convolution layers.

5. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 1 is characterized in that: Step S2 is specifically as follows: S21: obtaining a reference reservoir model TI and a reservoir model to be reconstructed Z, the size of the reservoir model to be reconstructed Z is H×L×W, and initializing the size of the conditional data to H×1×1; S22: partition the reservoir model Z to be reconstructed to obtain multiple regions, region Z i The dimensions are H'×L'×W'; S23: The region set of conditional data in the reservoir model Z to be reconstructed is Z ie ; S24: Divide the reference reservoir model TI into a first training set T1 of size H×L×W and a second training set T2 of size H'×L'×W'; S25: Train the initial MAE network using the first training set T1 to obtain a network model M1 of size H'×L'×W' with a mask rate of m 1, , calculate and obtain the loss function I1; S26: Train the initial MAE network through the second training set T2 to obtain a network model M2 with a size of H'×1×1, a mask rate of m2, and calculate the loss function I2; S27: Reconstruct the region set Z through conditional data and network model M2 ie The area where the number of pixels reaches the mask rate m1 required by the network model M1, and the reconstructed area set Z is obtained ie’ ; S28: According to the position of the conditional data in the reservoir model Z to be reconstructed, the reconstruction area set Z ie’ Restore to the position where it should be; S29: Reconstruct the region set Z ie’ As new condition data, the region set Z will be reconstructed ie’ Input the network model M1 to obtain the reconstructed reservoir model; S210: Repeat steps S22-S29 until the loss function I1 and the loss function I2 are both less than the preset values, and use the final network model M1 as the large-scale MAE network.

6. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 5 is characterized in that: The formula for loss function I1 is calculated as: Among them, λ1 is the first weight parameter, λ2 is the second weight parameter, A1 is the first activation matrix, It is the loss between the reconstructed reservoir model output by the network model M1 and the labeled reservoir model.

7. The large-scale reservoir model reconstruction method based on Transformer-CNN architecture according to claim 5 is characterized in that: The formula for loss function I2 is calculated as: Among them, λ3 is the third weight parameter, λ4 is the fourth weight parameter, A2 is the second activation matrix, It is the loss between the reconstructed reservoir model output by the network model M2 and the labeled reservoir model.

8. A large-scale reservoir model reconstruction device based on Transformer-CNN architecture, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the large-scale reservoir model reconstruction method based on Transformer-CNN architecture as described in any one of claims 1 to 7.

Images