Oil reservoir model global scale upgrading method and system based on deep learning

Abstract

The invention relates to an oil reservoir model global scale upgrading method and system based on deep learning, and the method comprises the steps: collecting fine scale information corresponding to a to-be-upgraded related parameter, obtaining input data, carrying out the scale upgrading numerical calculation of the input data, and obtaining an upscaling parameter as output data; correspondingly performing logarithmic transformation on the input data and the output data to construct a data set; training a deep learning model comprising a plurality of deep learning sub-models through the data set to obtain a deep learning upgrading model; and inputting the fine-scale information into the deep learning upgrading model to obtain predicted upscaling parameters. According to the method, a numerical calculation process in traditional scale upgrading is replaced by the deep learning model, the upscaling parameters required by the coarse scale model can be efficiently predicted in real time based on the initial information of the fine scale model, and time and computing resources are remarkably saved.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir modeling and numerical simulation, and in particular to a method and system for global scale upgrading of a reservoir model based on deep learning. Background Art

[0002] In order to accurately describe the strata, the fine geological model usually contains hundreds of millions of grids, and its numerical simulation requires huge computing resources and time costs. In order to ensure the high accuracy and efficiency of numerical simulation, it is necessary to carry out upscaling research to convert the fine-scale geological model into an accurate coarse-scale reservoir numerical simulation model. Among the existing methods, the global numerical method has the highest upscaling accuracy. However, this method usually involves complex numerical calculation and solution processes, with low computational efficiency and long time consumption.

[0003] Therefore, its application scope is usually limited to scenarios that require high accuracy of coarse-scale models or need to reuse coarse-scale models frequently. To address this problem, a method that can significantly improve the efficiency of global scale upscaling is urgently needed to broaden its practical application scope. Summary of the invention

[0004] The present invention provides a global scale upgrading method and system of a reservoir model based on deep learning, so as to solve the defects of the prior art.

[0005] The present invention provides a global scale upgrade method of a reservoir model based on deep learning, comprising:

[0006] S1: Collect fine-scale information corresponding to relevant parameters to be upgraded, obtain input data, perform numerical calculation on the scale upgrade of the input data, and obtain the scale upgrade parameters as output data;

[0007] S2: Perform logarithmic transformation on the input data and the output data to construct a data set;

[0008] S3: training a deep learning model including a plurality of deep learning sub-models through the data set to obtain a deep learning upgrade model;

[0009] S4: Inputting the fine-scale information into the deep learning upscaling model to obtain predicted upscaling parameters.

[0010] According to a method for global scale upgrading of a reservoir model based on deep learning provided by the present invention, after step S1, the method further includes:

[0011] Data cleaning is performed on the parameters with deviations in the output data, wherein the parameters with deviations specifically include abnormal values, outliers, and noise values.

[0012] According to a global scale upgrade method of a reservoir model based on deep learning provided by the present invention, in step S2, the input data and the output data are simultaneously logarithmically transformed, and the specific expression of the logarithmic transformation is:

[0013] ln(x+1*10 -6 );

[0014] Where x is the data to be logarithmically transformed.

[0015] According to a method for global scale upgrading of a reservoir model based on deep learning provided by the present invention, the deep learning model in step S3 includes:

[0016] a first deep learning sub-model for well index upscaling, the first deep learning sub-model taking the local permeability field, the wellbore location and the fine-scale well index as input and taking the upscaled well index as output;

[0017] a second deep learning sub-model for permeability upscaling, the second deep learning sub-model taking as input the local permeability field, the coarse-scale interface position, and the upscaled well index of the coarse grid upstream of the coarse-scale interface, and taking as output the upscaled conductivity;

[0018] a third deep learning sub-model for upscaling the capillary force curve, wherein the third deep learning sub-model takes the local permeability as input and takes the upscaled capillary force curve as output;

[0019] A fourth deep learning sub-model for relative permeability upscaling, wherein the fourth deep learning sub-model takes local permeability field, coarse-scale interface position, upscaled capillary force curve, upscaled well index and upscaled conductivity as input, and takes upscaled relative permeability as output.

[0020] According to a global scale upgrade method of a reservoir model based on deep learning provided by the present invention, the method of outputting the scaled well index by the first deep learning sub-model specifically includes:

[0021] S311: extracting features of the local permeability field through a CNN network to obtain a first local permeability field tensor;

[0022] S312: Processing the wellbore position and the fine-scale well index respectively through two fully connected layers to obtain a wellbore position tensor and a fine-scale well index tensor;

[0023] S313: Processing the fused tensor of the local permeability field tensor, the wellbore position tensor, and the fine-scale well index tensor through a Transformer encoder to obtain an upscaled well index.

[0024] According to a global scale upscaling method of a reservoir model based on deep learning provided by the present invention, the method of outputting the upscaled conductivity of the second deep learning sub-model specifically includes:

[0025] S321: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor;

[0026] S322: processing the upscaled well index of the coarse grid upstream of the coarse-scale interface and the coarse-scale interface position respectively through two fully connected layers to obtain an upscaled well index tensor and a coarse-scale interface position tensor;

[0027] S323: Processing the fused tensor of the local permeability field tensor, the upscaled well index tensor, and the coarse-scale interface position tensor through a Transformer encoder to obtain the upscaled conductivity.

[0028] According to a global scale upgrade method of a reservoir model based on deep learning provided by the present invention, the method of outputting the upscaled capillary force curve by the third deep learning sub-model specifically includes:

[0029] S331: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor;

[0030] S332: Transform the local permeability field tensor into an upscaled capillary force curve through multiple fully connected layers.

[0031] According to a global scale upgrade method of a reservoir model based on deep learning provided by the present invention, the method of outputting the upscaled permeability curve by the fourth deep learning sub-model specifically includes:

[0032] S341: extracting features of the local permeability field through CNN to obtain a local permeability field tensor;

[0033] S342: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a first tensor including a plurality of sub-tensors, and fusing the local permeability field tensor with the first tensor to obtain a first fused tensor;

[0034] S343: Convert the first fused tensor into a Fourier tensor through FNO and a fully connected layer;

[0035] S344: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a second tensor including a plurality of sub-tensors, and fusing the second tensor with the Fourier tensor to obtain a second fused tensor;

[0036] S345: Input the second fused tensor into the Transformer encoder for processing to obtain the upscaled relative permeability.

[0037] According to a deep learning-based global scale upgrading method for a reservoir model provided by the present invention, when the upscaling parameters included in the deep learning sub-model are anisotropic, the deep learning sub-models corresponding to the anisotropic upscaling parameters are trained respectively.

[0038] The present invention also provides a global scale upgrade system of a reservoir model based on deep learning, which is used to execute a global scale upgrade method of a reservoir model based on deep learning as described in any one of the above items, comprising:

[0039] Data collection module: used to collect fine-scale information corresponding to the relevant parameters to be upgraded, obtain input data, perform numerical calculations on the scale upgrade of the input data, and obtain the scale upgrade parameters as output data;

[0040] Data processing module: used to perform logarithmic transformation on the input data and the output data to construct a data set;

[0041] Training module: used for training a deep learning model including a plurality of deep learning sub-models through the data set to obtain a deep learning upgrade model;

[0042] Deep learning upgrade module: used to predict and obtain upscaling parameters through the deep learning upgrade model.

[0043] The present invention provides a global scale upgrade method and system for reservoir models based on deep learning. On the basis of the global scale upgrade method, a variety of deep learning technologies such as convolutional neural network (CNN), Transformer and Fourier neural operator (FNO) are combined to establish deep learning sub-models for upscaling well index, permeability, capillary force curve and relative permeability curve, and integrate these sub-models to form a complete global scale upgrade solution. The present invention effectively avoids the complex numerical calculation process in the traditional global scale upgrade method and greatly improves the scale upgrade efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1A schematic flow chart of a global scale upgrade method for a reservoir model based on deep learning provided by the present invention;

[0046] Figure 2 A schematic diagram of the structure of a global scale upgrade system for a reservoir model based on deep learning provided by the present invention;

[0047] Figure 3 A schematic diagram of the structure of the first deep learning sub-model provided by the present invention;

[0048] Figure 4 A schematic diagram of the structure of the second deep learning sub-model provided by the present invention;

[0049] Figure 5 A schematic diagram of the structure of the third deep learning sub-model provided by the present invention;

[0050] Figure 6 A schematic diagram of the structure of the fourth deep learning sub-model provided by the present invention;

[0051] Figure 7 A schematic diagram of upscaling parameter prediction samples with P90 MSE error values ​​in the validation set provided by an embodiment of the present invention;

[0052] Figure 8 A sample schematic diagram of gas phase flow relative errors of P90, P75, P50, and P25 in the test set provided by an embodiment of the present invention;

[0053] Fig. 9 A schematic diagram of a sample with a relative error of P90 for the gas saturation field in a test set provided by an embodiment of the present invention;

[0054] Fig.10 A schematic diagram of a sample with a relative error of P50 of the gas saturation field in the test set provided by an embodiment of the present invention;

[0055] Fig.11 A sample schematic diagram of a relative error of the pressure field in the test set of the embodiment of the present invention is P90.

[0056] Reference numerals:

[0057] 100. Data collection module; 200. Data processing module; 300. Training module; 400. Deep learning upgrade module. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0059] The embodiments of the present invention are described below with reference to the accompanying drawings.

[0060] like Figure 1 As shown, the present invention provides a global scale upgrade method of a reservoir model based on deep learning, comprising:

[0061] S1: Collect fine-scale information corresponding to relevant parameters to be upgraded, obtain input data, perform numerical calculations on the scale of the input data, and obtain scaled parameters as output data.

[0062] In step S1, first, according to the single-phase flow scaling method based on numerical method, including the well index scaling method, the permeability scaling method, and the two-phase flow scaling method, including the capillary force curve scaling method, the relative permeability scaling method, the corresponding calculations of the input and output quantities of the multiple deep learning sub-models in step S3 are performed respectively.

[0063] Wherein, after step S1, the method further includes:

[0064] Data cleaning is performed on parameters with deviations in the upscaled output data, wherein the parameters with deviations specifically include abnormal values, outliers, and noise values.

[0065] Furthermore, the above-mentioned outliers refer to values ​​with obvious errors, such as infinite values, Nan, negative values, relative permeability greater than 1, abnormal 0 or negative values, etc. Outliers are relatively easy to identify, and are usually replaced by calculating replacement values ​​during the scale-up process. However, since the replacement values ​​do not represent the accurate scale-up calculation results, all replacement values ​​should be excluded when constructing the training set of the deep learning model.

[0066] The above-mentioned outliers refer to values ​​whose calculated upscaling parameters are greatly different from normal values ​​but do not show the above-mentioned obvious errors. The present invention uses the isolation forest algorithm to screen out outliers.

[0067] The noise value mentioned above refers to the fact that there is a certain error in the upscaling parameter, but it is mixed with the normal value and does not appear outliers. The noise value is difficult to accurately identify, and it is necessary to improve the generalization ability of the deep learning model to weaken the impact of the noise value, such as using an algorithm that is insensitive to noise, increasing the number of samples in the training set, and strictly avoiding overfitting during the training process.

[0068] S2: Perform logarithmic transformation on the input data and the output data to construct a data set.

[0069] Wherein, in step S2, the input data and the output data are simultaneously logarithmically transformed, and the specific expression of the logarithmic transformation is:

[0070] ln(x+1*10 -6 );

[0071] Where x is the data to be logarithmically transformed.

[0072] In the training process of deep learning models, it is often difficult to capture small values ​​in the data set, which may lead to increased model training errors. In order to solve this problem, the present invention uses logarithmic transformation on both input and output data in the data set, i.e., ln(x+1*10 -6 ) transformation, this method effectively amplifies the differences in small values ​​in the data, making it easier for deep learning models to recognize and learn these differences, thereby improving training efficiency and effectiveness.

[0073] S3: Training a deep learning model including a plurality of deep learning sub-models through the data set to obtain a deep learning upgrade model.

[0074] Wherein, the deep learning model in step S3 includes:

[0075] A first deep learning sub-model for well index scale upscaling, wherein the first deep learning sub-model takes the local permeability field, wellbore location and fine-scale well index as input and takes the scaled well index as output.

[0076] The present invention deeply studies the mechanism of well index scale upgrading and analyzes its key influencing factors, takes the local permeability field, wellbore position and fine-scale well index as input features, and upgrades the well index WI * As an output, a deep learning sub-model for well exponential scaling was constructed, namely the first deep learning sub-model, which combines the neural network structure of convolutional neural network (CNN) and Transformer.

[0077] CNN performs well in processing data with grid structures and is particularly suitable for identifying and characterizing local permeability fields in this study. Through its unique structural design such as convolution and pooling, CNN can effectively identify and utilize the spatial hierarchy of data, thereby accurately capturing higher-level details and features. However, CNN mainly captures local features of data through convolution kernels, and has the problem of "small field of view", so it has certain limitations in understanding the overall structure of data and capturing long-distance dependencies.

[0078] To overcome this limitation, the present invention further adopts the Transformer model, which optimizes the ability to capture long-distance dependencies through the self-attention mechanism. Unlike CNN, which can only perceive information through convolution kernels, the Transformer can establish a direct connection between any two points in the entire sequence, which enables the model to better understand the overall structure of the data. In addition, this mechanism of the Transformer also enables parallel processing, which speeds up the efficiency of model training.

[0079] Therefore, by combining the local feature extraction advantages of CNN and the global correlation capture ability of Transformer, the constructed deep learning model can more comprehensively analyze the complex characteristics of the upscaling process, provide a deeper perspective for understanding and predicting upscaling parameters, and thus achieve significant improvements in the accuracy of deep learning predictions.

[0080] The method for the first deep learning sub-model to output the upscaled well index specifically includes:

[0081] S311: extracting features of the local permeability field through a CNN network to obtain a first local permeability field tensor;

[0082] S312: Processing the wellbore position and the fine-scale well index respectively through two fully connected layers to obtain a wellbore position tensor and a fine-scale well index tensor;

[0083] S313: Processing the fused tensor of the local permeability field tensor, the wellbore position tensor, and the fine-scale well index tensor through a Transformer encoder to obtain an upscaled well index.

[0084] Figure 3 This is a schematic diagram of the structure of the deep learning sub-model for well index scale upgrade proposed in the present invention. The model takes the local permeability field, wellbore position and fine-scale well index WI as input, and uses the scaled well index WI as input. *In this figure, the number before the @ symbol represents the number of channels, and the number after the @ symbol represents the tensor dimension of each channel. For example, 32@10×10 means that the output of this layer has 32 channels, each of which has a dimension of 10×10, while 1@128 means that the output of this layer is a single channel with a dimension of 128. **_n is the neural layer type and its corresponding number n. For example, Con_1 is a convolutional layer numbered 1, and FC_4 is a fully connected layer numbered 4.

[0085] like Figure 3 As shown in the figure, the local permeability field is first processed by CNN to identify important information and extract key features. The processed output tensor is 64 channels, each with a dimension of 2×2. The tensor is then flattened and converted to a 1@128 tensor through a fully connected layer. At the same time, the wellbore position and fine-scale well index are also processed by the fully connected layer and mapped into 1@128 tensors. These three 1@128 tensors are merged into a 3@128 tensor and then fed into the Transformer encoder.

[0086] The Transformer encoder integrates different input source information through the self-attention mechanism, analyzes and understands the inherent interactions and dependencies between them, and finally outputs a 1@1 upscaling well index WI * The Transformer encoder uses a total of 3 layers of encoders, 4 attention heads, and a feedforward neural network dimension of 256.

[0087] The second deep learning sub-model for permeability scaling takes the local permeability field, the coarse-scale interface position and the upscaled well index of the coarse grid upstream of the coarse-scale interface as input, and takes the upscaled conductivity as output.

[0088] Furthermore, both permeability scaling and well index scaling rely on global single-phase flow numerical simulation, and their calculation processes have certain similarities. Therefore, a neural network structure combining CNN and Transformer similar to the above is adopted.

[0089] By deeply exploring and analyzing the key influencing factors of permeability scaling, the present invention combines the local permeability field, the coarse-scale interface position, and the upscaling well index WI of the coarse grid upstream of the coarse-scale interface. * As input features (if there is no well in the upstream coarse grid of the coarse-scale interface, WI * is set to 0), and the upscaled conductivity T * as output.

[0090] The method for the second deep learning sub-model to output upscaled conductivity specifically includes:

[0091] S321: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor;

[0092] S322: processing the upscaled well index of the coarse grid upstream of the coarse-scale interface and the coarse-scale interface position respectively through two fully connected layers to obtain an upscaled well index tensor and a coarse-scale interface position tensor;

[0093] S323: Processing the fused tensor of the local permeability field tensor, the upscaled well index tensor, and the coarse-scale interface position tensor through a Transformer encoder to obtain the upscaled conductivity.

[0094] Figure 4 Taking the x-direction flow as an example, the deep learning sub-model structure of permeability upscaling is demonstrated. * Defined on the coarse-scale interface, the input local permeability field contains two coarse grids adjacent to the coarse-scale interface, with a size of 20 × 10. The Transformer encoder used has the same structure as the Transformer encoder in the first deep learning sub-model mentioned above.

[0095] A third deep learning sub-model for upscaling capillary force curves, wherein the third deep learning sub-model takes local permeability as input and takes an upscaled capillary force curve as output.

[0096] The present invention defines the fine-scale capillary force curve based on the Leverett J function, and its calculation formula is as follows:

[0097]

[0098] Where k is the grid permeability, k ref is the reference permeability, φ is the grid porosity, φ ref is the reference porosity, J(S g ) is the Leverett J function, S g is the gas saturation.

[0099] The present invention adopts a network structure that combines CNN and a fully connected layer. Based on the homogeneous porosity condition assumed in this embodiment, the capillary force curve is only related to the permeability. Therefore, the local permeability is used as the input feature of the model, and the upscaled capillary force curve is used as the output for training the deep learning model, which effectively improves the efficiency of the capillary force curve scale upgrade.

[0100] The method for the third deep learning sub-model to output the upscaled capillary force curve specifically includes:

[0101] S331: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor;

[0102] S332: Transform the local permeability field tensor into an upscaled capillary force curve through multiple fully connected layers.

[0103] Figure 5 The structure of the deep learning sub-model for upscaling the capillary force curve is presented. First, the local permeability field is input into CNN to extract its key features, and a tensor with a dimension of 64@2×2 is obtained. After the tensor is flattened, it is passed through 4 layers of fully connected layers in succession, and finally an upscaled capillary force curve with a dimension of 1@12 is obtained.

[0104] A fourth deep learning sub-model for relative permeability upscaling, wherein the fourth deep learning sub-model takes local permeability field, coarse-scale interface position, upscaled capillary force curve, upscaled well index and upscaled conductivity as input, and takes upscaled relative permeability as output.

[0105] Furthermore, the upscaling of relative permeability requires the pre-execution of global fine-scale two-phase flow numerical simulation, which is the most complex part of the entire upscaling process. Because it involves the solution of complex two-phase flow partial differential equations and the upscaling calculation process, simple deep learning algorithms are difficult to accurately capture its key features. In addition, since relative permeability is a dynamic property that changes with saturation, the difficulty of deep learning models in capturing this continuously changing feature is significantly greater than that of single-phase flow upscaling. In addition, the complex calculation process results in more noise values ​​generated during the upscaling calculation process than single-phase flow upscaling, which brings additional challenges to deep learning models.

[0106] In order to effectively capture the complex features in the process of solving the two-phase flow partial differential equations, the present invention introduces the Fourier neural operator (FNO) in the deep learning model of relative permeability scale upgrade. FNO shows significant advantages in dealing with partial differential equation problems. It can use Fourier transform to convert space-time domain information into the frequency domain (i.e., Fourier space), thereby effectively capturing and processing information and features that are difficult to reveal in the space-time domain. By analyzing the frequency component information in Fourier space, FNO can not only capture the global characteristics and continuity information of partial differential equations, but also reveal the long-distance dependencies therein. Therefore, FNO can extract richer information from the training samples, and then learn and understand the physical processes and laws behind the samples, which is particularly important for in-depth understanding and characterization of the complex physical mechanisms in the two-phase flow process.

[0107] The method for the fourth deep learning sub-model to output an upscaled permeability curve specifically includes:

[0108] S341: extracting features of the local permeability field through CNN to obtain a local permeability field tensor;

[0109] S342: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a first tensor including a plurality of sub-tensors, and fusing the local permeability field tensor with the first tensor to obtain a first fused tensor;

[0110] S343: Convert the first fused tensor into a Fourier tensor through FNO and a fully connected layer;

[0111] S344: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a second tensor including a plurality of sub-tensors, and fusing the second tensor with the Fourier tensor to obtain a second fused tensor;

[0112] S345: Input the second fused tensor into the Transformer encoder for processing to obtain the upscaled relative permeability.

[0113] Taking the x direction as an example, the structure of the deep learning sub-model for relative permeability scaling is as follows: Figure 6 As shown in the figure, the model integrates CNN, Transformer and FNO. CNN can effectively extract the characteristics of the local permeability field, providing a solid feature representation foundation for the subsequent processing of the model. FNO can mine richer information from the data and effectively learn the global continuity characteristics, thereby improving the model's ability to understand complex flows. Transformer can effectively capture long-distance dependencies through the self-attention mechanism, analyze the global correlation between various inputs, and enhance the model's ability to process global information, thereby achieving a more accurate understanding and prediction of two-phase flow behavior. The detailed description of the model structure is as follows:

[0114] The specific processing flow is as follows: first, the local permeability field is input into CNN, and key features are extracted through operations such as convolution and pooling, and represented as a 128@3×1 tensor. Then, other input features (including coarse-scale interface position, upscale capillary force, upscale well index, and upscale conductivity) are mapped into 1@16 tensors through fully connected layers, and the output tensor of CNN is flattened and combined with it to become a 1@448 tensor. The tensor is further converted into a 1@512 tensor through a fully connected layer and then input into FNO. FNO consists of 3 layers. The first layer has 1 input channel and 128 output channels; the last two layers have 128 input and output channels. The output of FNO is a 128@512 tensor, which is converted into a 1@128 tensor after being processed by the fully connected layer. Finally, other input features (including coarse-scale interface position, upscale capillary force, upscale well index and upscale conductivity) are mapped into a 1@128 tensor through a fully connected layer again, and combined with the tensor output by FNO to form a tensor of dimension 5@128, which is then input into the Transformer encoder for regression analysis. The Transformer encoder used has 3 layers, 8 attention heads, and a feedforward neural network dimension of 256. Finally, the Transformer outputs an upscaled relative permeability of 1@12.

[0115] Among them, when the upscaling parameters included in the deep learning sub-model are anisotropic, the deep learning sub-models corresponding to the anisotropic upscaling parameters are trained separately, that is, for the gas and water phase models in different directions, the second deep learning sub-model and the fourth deep learning sub-model both include corresponding models in the x-direction and the z-direction. However, it should be noted that when training the z-direction model, it is only necessary to adjust the input local permeability field size to 10×20, the model structure does not need to be changed, and the final output is not affected.

[0116] In addition, during the model training process in step S3, the deep learning sub-models of each part are trained separately, the batch size of the training is set to 256, and the initial learning rate is 10 -4 , the optimizer uses Adam, and the loss function uses mean square error (MSE) loss.

[0117] S4: Obtain upscaling parameters through prediction of the deep learning upscaling model.

[0118] After steps S1 to S3, after completing the training of each part of the deep learning sub-model, the final step S4 is to perform prediction. In step S4, the present invention encapsulates and integrates the deep learning sub-models of each part, and only needs to input the permeability field and well information of the fine-scale model into the integrated model, and finally outputs the upscaling parameters required for the coarse-scale model.

[0119] like Figure 2 As shown, the present invention also provides a global scale upgrade system of a reservoir model based on deep learning, comprising:

[0120] Data collection module 100: used to collect fine-scale information corresponding to the relevant parameters to be upgraded, obtain input data, perform numerical calculation of the scale upgrade on the input data, and obtain the scale upgrade parameters as output data;

[0121] Data processing module 200: used to perform logarithmic transformation on the input data and the output data to construct a data set;

[0122] Training module 300: used to train a deep learning model including multiple deep learning sub-models through the data set to obtain a deep learning upgrade model;

[0123] The deep learning upgrade module 400 is used to predict and obtain the upscaling parameters through the deep learning upgrade model.

[0124] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0125] The embodiments of the present invention are described below in conjunction with specific examples.

[0126] This example takes CO2 saline storage as an example. The fine-scale model contains a grid number of 200×100, which is converted into a coarse-scale model of 20×10 after scale up. The model is defined on the xz two-dimensional coordinate axis, with initial saturated brine, a row of CO2 injection wells on the left boundary of the model, and a constant pressure boundary on the right boundary, with the fluid flowing from left to right.

[0127] This embodiment randomly generates 8,000 geological models, and constructs training sets, validation sets, and test sets by random division in the proportion of 40%, 10%, and 50%, respectively. First, the prediction accuracy of the deep learning model for the upscaling parameters in the validation set is evaluated. Then, the coarse-scale numerical simulation results of the upscaling model based on deep learning in the test set are compared with the numerical simulation results of the fine-scale model as a reference. The deep learning model provided is written in Python and runs on a Tesla A100 80G GPU.

[0128] Figure 7 Shows the validation set with P90 MSE error values The results show that the upscaling parameters predicted by the deep learning upscaling model of the present invention are very accurate. Since the figure shows the P90 error, the MSE error of 90% of the samples in the validation set is lower than the result shown in the figure.

[0129] Figure 8 The figure shows the samples of gas phase flow rate relative errors of P90, P75, P50, and P25 in the upscaling model based on deep learning. The figure shows that in the coarse-scale model predicted by deep learning, 90% of the gas phase flow rate is Figure 8 The result shown in a is more accurate, with 75% of the gas phase flow rate being Figure 8 The result shown in b is more accurate, with a gas flow rate of 50% Figure 8 c is more accurate, 25% of the gas phase flow rate is Figure 8 The result shown in (d) is more accurate. The analysis results show that even for samples with a relative error of gas phase flow reaching P90, the accuracy is relatively ideal. This shows that upscaling based on deep learning can provide fairly accurate coarse-scale gas phase flow results for most models.

[0130] Fig. 9 and Fig.10 The samples with relative errors of gas saturation of P90 and P50 in the upscaling model based on deep learning are shown respectively. Under the conditions of PVI of 0.1 (no gas at the constant pressure boundary) and 0.5 (gas at the constant pressure boundary), the average fine-scale gas saturation field and the gas saturation field of the upscaling model based on deep learning are compared and analyzed. The results shown in the figure are relatively accurate, the gas saturation of the coarse-scale model is highly consistent with the fine-scale model, and the gas distribution is roughly the same. Since the figure shows the P90 and P50 errors respectively, the gas saturation field accuracy of 90% of the coarse-scale models in the test set is higher than Fig. 9 The results show that the gas saturation field accuracy of the 50% coarse-scale model is higher than Fig.10 This shows that the deep learning model can capture complex seepage characteristics and provide relatively accurate gas saturation field results for most coarse-scale models.

[0131] In the accuracy assessment of the pressure field, the deep learning-based upscaling model showed very high accuracy. Fig.11 This is a sample with a relative error of P90 for the pressure field in the deep learning-based upscaling model. The results show that the pressure field accuracy of the deep learning-based upscaling model is very high. The pressure field of 90% of the coarse-scale models predicted by the deep learning model is more accurate than this figure. This proves that the vast majority of deep learning-based upscaling models can obtain very accurate pressure field results.

[0132] The comparison of the calculation time results of the upscaling method based on deep learning of the present invention with those of the traditional numerical method for upscaling is shown in Table 1. The use of the deep learning model instead of the traditional numerical method for upscaling calculation improves the upscaling efficiency by 1133 times.

[0133] Table 1 Comparison of the computation time of the scaling method of the present invention

[0134] Scale Up Method Calculation time / s Speedup Traditional numerical method scaling up 1700 — A deep learning-based upscaling approach 1.5 1133

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A global scale upgrade method for reservoir models based on deep learning, characterized in that: include: S1: Collect fine-scale information corresponding to relevant parameters to be upgraded, obtain input data, perform numerical calculation on the scale upgrade of the input data, and obtain the scale upgrade parameters as output data; S2: Perform logarithmic transformation on the input data and the output data to construct a data set; S3: training a deep learning model including a plurality of deep learning sub-models through the data set to obtain a deep learning upgrade model; S4: Inputting the fine-scale information into the deep learning upscaling model to obtain predicted upscaling parameters.

2. The method for global scale upgrading of reservoir model based on deep learning according to claim 1, characterized in that: After step S1, the method further includes: Data cleaning is performed on the parameters with deviations in the output data, wherein the parameters with deviations specifically include abnormal values, outliers, and noise values.

3. The method for global scale upgrading of reservoir model based on deep learning according to claim 1, characterized in that: In step S2, the input data and the output data are simultaneously logarithmically transformed. The specific expression of the logarithmic transformation is: ln(x+1*10 -6 ); Where x is the data to be logarithmically transformed.

4. The method for global scale upgrading of reservoir model based on deep learning according to claim 1, characterized in that: The deep learning model in step S3 includes: a first deep learning sub-model for well index upscaling, the first deep learning sub-model taking the local permeability field, the wellbore location and the fine-scale well index as input and taking the upscaled well index as output; a second deep learning sub-model for permeability upscaling, the second deep learning sub-model taking as input the local permeability field, the coarse-scale interface position, and the upscaled well index of the coarse grid upstream of the coarse-scale interface, and taking as output the upscaled conductivity; a third deep learning sub-model for upscaling the capillary force curve, wherein the third deep learning sub-model takes the local permeability as input and takes the upscaled capillary force curve as output; A fourth deep learning sub-model for relative permeability upscaling, wherein the fourth deep learning sub-model takes local permeability field, coarse-scale interface position, upscaled capillary force curve, upscaled well index and upscaled conductivity as input, and takes upscaled relative permeability as output.

5. The method for global scale upgrading of reservoir model based on deep learning according to claim 4, characterized in that: The method for the first deep learning sub-model to output the upscaled well index specifically includes: S311: extracting features of the local permeability field through a CNN network to obtain a first local permeability field tensor; S312: Processing the wellbore position and the fine-scale well index respectively through two fully connected layers to obtain a wellbore position tensor and a fine-scale well index tensor; S313: Processing the fused tensor of the local permeability field tensor, the wellbore position tensor, and the fine-scale well index tensor through a Transformer encoder to obtain an upscaled well index.

6. The method for global scale upgrading of reservoir model based on deep learning according to claim 4, characterized in that: The method for the second deep learning sub-model to output upscaled conductivity specifically includes: S321: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor; S322: processing the upscaled well index of the coarse grid upstream of the coarse-scale interface and the coarse-scale interface position respectively through two fully connected layers to obtain an upscaled well index tensor and a coarse-scale interface position tensor; S323: Processing the fused tensor of the local permeability field tensor, the upscaled well index tensor, and the coarse-scale interface position tensor through a Transformer encoder to obtain the upscaled conductivity.

7. The method for global scale upgrading of reservoir model based on deep learning according to claim 4, characterized in that: The method for the third deep learning sub-model to output the upscaled capillary force curve specifically includes: S331: extracting features of the local permeability field through a CNN network to obtain a local permeability field tensor; S332: Transform the local permeability field tensor into an upscaled capillary force curve through multiple fully connected layers.

8. The method for global scale upgrading of reservoir model based on deep learning according to claim 4, characterized in that: The method for the fourth deep learning sub-model to output an upscaled permeability curve specifically includes: S341: extracting features of the local permeability field through CNN to obtain a local permeability field tensor; S342: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a first tensor including a plurality of sub-tensors, and fusing the local permeability field tensor with the first tensor to obtain a first fused tensor; S343: Convert the first fused tensor into a Fourier tensor through FNO and a fully connected layer; S344: mapping the coarse-scale interface position, the upscaled well index, the upscaled conductivity, and the upscaled capillary force curve respectively through a fully connected layer to obtain a second tensor including a plurality of sub-tensors, and fusing the second tensor with the Fourier tensor to obtain a second fused tensor; S345: Input the second fused tensor into the Transformer encoder for processing to obtain the upscaled relative permeability.

9. The method for global scale upgrading of reservoir model based on deep learning according to claim 4, characterized in that: When the upscaling parameters included in the deep learning sub-model are anisotropic, the deep learning sub-models corresponding to the anisotropic upscaling parameters are trained separately.

10. A global scale upgrade system for a reservoir model based on deep learning, used to execute a global scale upgrade method for a reservoir model based on deep learning as claimed in any one of claims 1 to 9, characterized in that: include: Data collection module: used to collect fine-scale information corresponding to the relevant parameters to be upgraded, obtain input data, perform numerical calculations on the scale upgrade of the input data, and obtain the scale upgrade parameters as output data; Data processing module: used to perform logarithmic transformation on the input data and the output data to construct a data set; Training module: used for training a deep learning model including a plurality of deep learning sub-models through the data set to obtain a deep learning upgrade model; Deep learning upgrade module: used to predict and obtain upscaling parameters through the deep learning upgrade model.

Images