A dynamic optimization method for coal seam hydraulic fracturing parameters based on deep learning

Abstract

The present invention provides a dynamic optimization method for coal seam hydraulic fracturing parameters based on deep learning, belonging to the technical field of coal seam parameter optimization based on computer data processing; First, collect CT images of coal seam slices with spatial continuity; Secondly, design an intelligent perception model for coal seam structure-physical parameters to jointly predict semantic recognition and physical property estimation of coal seam images; Subsequently, use the predicted output results to reconstruct the three-dimensional structure of the coal seam and build a regional stratification model, and on this basis, conduct finite element mesh division and accurate mapping of material properties to construct a three-dimensional finite element model of the coal seam; Finally, construct a multi-objective optimization problem for coal seam hydraulic fracturing parameters, design a surrogate model embedded with coal seam physical information and combine it with a multi-objective evolutionary algorithm to dynamically solve the optimization problem and obtain the optimal hydraulic fracturing parameters. The present invention is based on deep learning, integrates multi-source information, structural parameters and material properties, and promotes the intelligent development of coal seam hydraulic fracturing design.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal seam parameter optimization based on computer data processing, and particularly relates to a dynamic optimization method for coal seam hydraulic fracturing parameters based on deep learning. Background Art

[0002] With the continuous advancement of coalbed methane resource development, coal seam hydraulic fracturing, as a key means to improve coal seam permeability and enhance coalbed methane production capacity, has become one of the core technologies in the process of coal mining and coalbed methane development. Hydraulic fracturing forms a fracture network in the coal seam by injecting fracture fluid at high pressure, thereby enhancing the conductivity of the coal seam. The setting of fracturing parameters, such as injection pressure, displacement, proppant concentration, and injection time, directly determines the fracturing effect and economic benefits. However, due to the complex coal seam structure and variable physical parameters, traditional fracturing designs usually rely on experience or simplified models and are difficult to dynamically respond to on-site geological changes, affecting the fracture development effect and gas production stability. In recent years, the development of artificial intelligence and deep learning technologies has provided new possibilities for the intelligent optimization of coal seam hydraulic fracturing parameters. Through in-depth analysis of on-site images and material properties, more accurate prediction and control of coal seam fracturing behavior can be achieved, providing a basis for formulating more scientific fracturing parameters.

[0003] The existing methods for formulating coal seam hydraulic fracturing parameters mainly include the following three types:

[0004] Fracturing design method based on on-site experience and static geological parameters: This method usually manually formulates a fracturing parameter scheme by engineers according to geological exploration data (such as coal seam thickness, in-situ stress, fracture density, etc.) and past experience. The common processes include geological parameter investigation, selection of experience templates, and manual adjustment of parameters. Such methods were relatively common in early coal seam fracturing projects and had certain practicability. However, such methods rely heavily on the experience of construction personnel and are difficult to adapt to complex or atypical geological conditions. In areas with strong coal seam structure heterogeneity, this method often cannot accurately reflect the influence of microscopic structure on fracturing response, and there are problems such as large parameter setting deviations and unstable fracturing effects;

[0005] Numerical simulation method based on single physical property input: This method constructs a finite element model through structural parameters such as in-situ stress, porosity, and permeability, and uses fracturing simulation software (such as ABAQUS, FLAC3D) to predict the fracture propagation path and influence range, and iteratively optimizes the fracturing parameters based on the simulation. However, model construction relies on simplified expressions of geological parameters, the input data dimension is low, and it is difficult to comprehensively describe the heterogeneity and detailed characteristics of the coal seam. In addition, this method usually ignores the influence of actual material forms (such as fracture distribution, parting layers, etc.) on the simulation results, resulting in limited simulation accuracy and a certain deviation in the guiding role for actual construction;

[0006] Fracturing effect prediction method based on machine learning model: This method collects historical fracturing construction data (including fracturing fluid consumption, injection rate, proppant strength, etc.) and corresponding coal seam characteristics, and uses traditional machine learning algorithms (such as decision trees, support vector machines, etc.) to establish a regression or classification model between fracturing parameters and gas production effects, so as to realize the prediction of fracturing effects and suggestions for parameter adjustment. However, such methods rely highly on data quality and sample size, and the model training process lacks an explanation of physical mechanisms. In addition, since the model input is mainly based on numerical variables and lacks the ability to model non-numerical features such as images and structures, its generalization performance is poor when facing new well areas or complex structural conditions.

[0007] Therefore, existing methods all have different degrees of limitations in the process of optimizing fracturing parameters, and it is difficult to achieve precise perception and dynamic response to complex coal seam structures. Summary of the Invention

[0008] In view of the above problems, the present invention proposes a dynamic optimization method for coal seam hydraulic fracturing parameters based on deep learning, including the following steps:

[0009] S1, collect CT images of coal seam slices with spatial continuity;

[0010] S2, input the CT images into the intelligent perception model of coal seam structure - physical parameters that has been constructed and trained. The model includes a coal seam multi-scale spatial attention feature encoder and a bedding context perception decoder, and is used to output the coal seam semantic segmentation result and corresponding material physical parameters;

[0011] S3, perform regional stratification and grid division according to the semantic segmentation result, and then map the corresponding material physical parameters to each unit of the finite element model according to the semantic region to obtain a three-dimensional finite element model of the coal seam;

[0012] S4, establish a multi-objective optimization problem for coal seam hydraulic fracturing parameters with the optimization objectives of maximizing the transformation volume, minimizing the comprehensive cost, and maximizing the fracture conductivity, and using the coal seam hydraulic fracturing parameters to be solved as decision variables; solve the multi-objective optimization problem based on the surrogate-assisted multi-objective evolutionary algorithm to obtain the optimal coal seam hydraulic fracturing parameters;

[0013] The surrogate-assisted multi-objective evolutionary algorithm is to generate training data and test data from the constructed three-dimensional finite element model of the coal seam to train and test the surrogate model of the coal seam physical information embedding to obtain a global surrogate model, and combine the global surrogate model with the multi-objective evolutionary algorithm to dynamically solve the multi-objective optimization problem.

[0014] Preferably, based on the training task of the model, collect and construct a coal seam semantic segmentation and material calibration data set, and the specific process is as follows:

[0015] First, for the coal seam body to be collected, obtain a sequence of slice CT images with spatial continuity: evenly divide the coal seam body into slices, and each group of CT image data contains slice images collected at equal intervals between layers, that is, , and a unified coordinate system calibration is adopted during the collection process;

[0016] Then, semantically annotate the key coal seam geological structures in the slice images, including 8 categories: background, main coal seam, interlayer parting, original fracture, fault / joint, mineralized zone, weak interlayer, and aquifer. Finally, obtain the semantic segmentation maps of the slice images ;

[0017] Finally, on the basis of semantic annotation, calibrate the corresponding material physical parameters for the regions of each structural category. The calibrated parameters include elastic modulus , Poisson's ratio , permeability , tensile strength , shear strength , fracture toughness , fracture-matrix coupling ratio , saturation , fracturing-induced fracture threshold , fracturing priority ; For the slice images, a total of groups of material physical parameter maps are obtained. The material physical parameter maps contain the material physical parameters corresponding to each coal seam semantic region.

[0018] Preferably, the specific processing process of the layer structure-physical parameter intelligent perception model is as follows:

[0019] S21. For the th slice CT image of each layer to be segmented , where ; First, use two 3*3 convolutional blocks to perform preliminary feature extraction on the CT image. Each 3*3 convolutional block consists of a convolutional layer with a convolution kernel of 3*3, a batch normalization layer, and a ReLU activation function, which are used to extract local texture information and basic spatial features. Subsequently, three max-pooling layers are sequentially introduced for multi-scale downsampling to obtain coal seam structure features at different spatial scales, and the feature outputs from shallow to deep are , corresponding to the front-layer features, middle-layer features, and deep-layer features respectively;

[0020] S22. The multi-scale features The input coal seam multi-scale spatial attention feature encoder realizes the extraction of deep features of the coal seam structure and the fusion of bedding context information:

[0021] After upsampling the deep features and fusing them with the middle-level features through the second hybrid feature encoding block, fused features are obtained; subsequently, continue to upsample and further fuse with the shallow features through the first hybrid feature encoding block to obtain the shallow fusion result ;

[0022] After average pooling the shallow fusion result and inputting it together with the fused features into the third hybrid feature encoding block to enhance the medium-scale feature perception ability and obtain the fused features ; subsequently, further average pool and input it together with the deep features into the fourth hybrid feature encoding block to output the fusion result ;

[0023] Concatenate the fused features , , of three different scales in channels to construct a multi-scale feature set; finally, enhance the spatial distribution characteristics of the key structures in the image through the spatial attention module, and finally output the bedding structure feature map representing the th slice image;

[0024] S23, based on the bedding context-aware decoder, performs image semantic reconstruction by combining the bedding information of adjacent slices:

[0025] For the bedding structure features of the th slice, concatenate the bedding structure features and of its two adjacent front and back slices in channels to form the context fusion features ; then, the context fusion features are input into a 3*3 transposed convolution block for preliminary decoding upsampling, and the multi-head attention mechanism is introduced to explicitly model and enhance the features of the key structures in the bedding context to obtain the enhanced features ;

[0026] Subsequently, Decode through four parallel feature reconstruction paths. Each path extracts information from different receptive fields and modeling strategies, and splices the outputs of the four paths in channels to obtain the feature map of the bedding structure of the th slice image layer Decoder output features ;

[0027] S24. Finally, the is compressed by the output layer and mapped respectively to obtain the estimated results of the final semantic segmentation map of the slice CT image and the estimated results of the material physical parameter map .

[0028] Preferably, the first hybrid feature encoding block, the second hybrid feature encoding block, the third hybrid feature encoding block, and the fourth hybrid feature encoding block are specifically:

[0029] For two groups of input features to be fused, first splice them in the channel dimension to form a joint feature representation; subsequently, the joint feature is input into three parallel paths simultaneously, extracting spatial and semantic information from different perspectives: Path 1 adopts a residual connection structure with batch normalization to retain the original feature trend and boundary information; Path 2 first compresses the channels through a 1*1 convolutional block and then extracts local structural features through a max-pooling operation to enhance the model's response ability to key regions of the main coal seam boundary and parting blocks; Path 3 performs channel mapping through a 1*1 convolutional block, then introduces a RepVGG layer with deep modeling ability to obtain a stronger semantic representation, and finally further improves the perception accuracy of spatial bedding changes through a 3*3 convolutional layer; finally, splice the features extracted from the three paths in channels to form a fused feature representation.

[0030] Preferably, the is decoded through four parallel feature reconstruction paths, specifically:

[0031] Among them, the first path combines 3*1 and 1*3 transposed convolution blocks to enhance the direction perception ability of the horizontal and vertical boundary structures of the bedding; the second path adopts 3*3 and 1*1 transposed convolution blocks to fuse local details and global contour information; the third path uses a combination of 5*5 and 1*1 transposed convolution blocks to expand the receptive field and capture larger-scale bedding change features; the fourth path effectively retains the original semantic features and strengthens the deep structure information through a residual connection mechanism.

[0032] Preferably, the specific process of S3 is:

[0033] S31, 3D structure reconstruction and regional stratification modeling: First, based on the estimated results of the semantic segmentation map of the sequence CT slice images Perform three-dimensional structure reconstruction; obtain the three-dimensional semantic structure distribution of the coal seam body by arranging equidistant slice images in sequence in a unified coordinate system, construct corresponding structural voxel models for different semantic categories, and perform regional stratification to obtain a stratified semantic structure model;

[0034] S32, Mesh generation and structure discretization: Based on the stratified semantic structure model, use the three-dimensional finite element method for structure discretization, adopt an unstructured mesh generation strategy for complex geological structure forms, and adaptively adjust the element density in combination with the geometric characteristics of different regions to finally obtain a discretized finite element model;

[0035] S33, Material physical parameter mapping: Map various parameters in the material physical parameter map estimation result accurately to the finite element mesh elements according to the semantic regions; this mapping process supports pixel-by-pixel parameter correspondence, enabling the model to finely reflect the heterogeneity and spatial variability of the physical properties inside the coal seam, and obtain a three-dimensional finite element model of the coal seam , denotes the material physical parameters.

[0036] Preferably, the establishment of the multi-objective optimization problem for coal seam hydraulic fracturing parameters in S4 is specifically as follows:

[0037] Determine the decision variables: Select 9 key parameters as decision variables, including 6 fracturing construction parameters and 3 engineering control parameters; among them, the fracturing construction parameters include the displacement , whose engineering significance is to control the crack propagation speed; the sand ratio , whose engineering significance is to affect the proppant transportation efficiency; the fracturing fluid viscosity , whose engineering significance is to affect the sand-carrying capacity and filtration loss; the injection time , whose engineering significance is to control the crack extension duration; the stage volume , used to control the scale of single-stage fracturing; the pumping rate , whose engineering significance is to affect the construction continuity; the engineering control parameters include the perforation cluster spacing , whose engineering significance is to control the crack initiation position; the proppant particle size , whose engineering significance is to balance the conductivity and settlement; the temporary plugging agent concentration , whose engineering significance is to affect the multi-cluster expansion balance;

[0038] Design the objective function:

[0039] ;

[0040] Among them is the transformation volume objective function, and minimizing this objective function is used to improve the utilization rate of natural fractures in the coal seam, is the modified volume, obtained by calculating the 3D finite element model of the coal seam; is the comprehensive cost objective function, covering the costs of proppant, fracturing fluid, and temporary plugging agent consumption. Minimizing this objective function is used to reduce the cost of hydraulic fracturing of the coal seam. ， ， and are the fixed cost coefficients of displacement, sand ratio, stage volume, and temporary plugging agent concentration respectively; is the fracture conductivity objective function. Minimizing this objective function is used to improve the fracture conductivity after hydraulic fracturing of the coal seam. is the conductivity coefficient, is the fracture permeability, obtained by calculating the 3D finite element model of the coal seam.

[0041] Preferably, the global surrogate model includes an encoder, a large-scale physical information extraction layer, a small-scale physical information extraction layer, a physical information aggregation layer, and a fully connected layer; the data data is input into the global surrogate model to obtain and predicted values. The specific steps are as follows:

[0042] The data is input into the encoder to extract the primary features of each decision variable data, obtaining the decision variable primary feature matrix ; the encoder consists of a convolutional layer and a RELU activation function layer. The convolutional layer is used to extract the local features of each decision variable data, and the RELU activation function layer is used to perform a non-linear transformation on the local features of each decision variable data;

[0043] The data is input into the large-scale physical information extraction layer to extract the large-scale correlation features of each decision variable data, obtaining the large-scale correlation feature matrix , and the large-scale physical information extraction layer uses a graph attention neural network with a convolutional kernel scale of to extract the large-scale features of each decision variable data in the data, and calculate the correlation coefficients for the large-scale features of each decision variable to obtain ;

[0044] The data is input into the small-scale physical information extraction layer to extract the small-scale correlation features of each decision variable data, obtaining the small-scale correlation feature matrix , and the large-scale physical information extraction layer uses a graph attention neural network with a convolutional kernel scale of to extract the small-scale features of each decision variable data in the data, and calculate the correlation coefficients for the small-scale features of each decision variable to obtain ;

[0045] Add and , then input the sum together with into the physical information aggregation layer to aggregate the graph features and local features to obtain aggregated features ; The physical information aggregation layer uses a graph aggregation neural network to sample and aggregate the decision variable features in each of the decision variable features in F and the feature map after adding and respectively, to obtain aggregated features representing the implicit relationship of decision variable data ;

[0046] Input into the fully connected layer to reduce the dimension and map the features to obtain and predicted values.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) A coal seam structure-physical parameter intelligent perception model integrating multi-scale spatial attention and bedding context modeling is proposed, which can simultaneously realize the semantic segmentation of coal seam images and the joint estimation of regional physical parameters, significantly improving the recognition accuracy of coal seam heterogeneous structures and the physical property perception ability;

[0049] (2) A three-dimensional finite element modeling method for coal seam complex structures and material heterogeneity is constructed. Based on the semantic segmentation results, regional stratification and mesh generation are carried out, and the material parameters output by the perception model are accurately mapped to the finite element units, realizing high-fidelity simulation modeling of coal seam structure-physical property integration;

[0050] (3) A fracturing parameter optimization strategy integrating a global surrogate model with coal seam physical information embedding and a multi-objective evolutionary algorithm is designed, which can efficiently approximate the global response law based on finite element simulation and dynamically solve the optimal fracturing plan that satisfies multiple constraints of fracturing effect, construction cost and diversion capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the overall technical route flow chart of the present invention.

[0052] Figure 2 is the overall framework diagram of the coal seam structure-physical parameter intelligent perception model.

[0053] Figure 3 is the schematic diagram of the structure of the surrogate model with coal seam physical information embedding.

[0054] Figure 4 is the experimental result diagram of coal seam structure semantic segmentation in the embodiment.

[0055] Figure 5 The hypervolume convergence curve in the embodiment.

[0056] Figure 6 The visualization of the optimal population objective value in the embodiment. Specific implementation manner

[0057] The present invention proposes a dynamic optimization method for coal seam hydraulic fracturing parameters based on deep learning. First, based on coal seam CT slice images, a semantic segmentation and calibration data set covering the coal seam geological structure and its material physical properties is constructed; second, an intelligent perception model for coal seam structure-physical parameters is designed to realize the joint prediction of semantic recognition and physical property estimation of coal seam images; subsequently, the output results of the intelligent perception model for coal seam structure-physical parameters are used for three-dimensional coal seam structure reconstruction and regional stratification modeling, and on this basis, finite element mesh division and accurate mapping of material properties are carried out to construct a three-dimensional finite element model of the coal seam that can truly represent the inhomogeneous structure and complex physical characteristics of the coal seam; finally, a multi-objective optimization problem for coal seam hydraulic fracturing parameters is constructed, and a surrogate model embedded with coal seam physical information is designed and combined with a multi-objective evolutionary algorithm to dynamically solve the optimization problem, and the optimal hydraulic fracturing engineering parameters that meet the balance of transformation effect, construction cost and diversion capacity are obtained.

[0058] The overall process of this embodiment is as Figure 1 shown:

[0059] Production of coal seam semantic segmentation and material calibration data set: First, systematically collect a group of coal seam slice CT images in groups, and perform semantic annotation on the key geological structures in the CT images (including background, main coal seam, interlayer gangue, original fractures, faults / joints, mineralized zones, soft interlayers, aquifers), and for each semantic region, calibrate its corresponding material physical parameters (including elastic modulus, Poisson's ratio, permeability, tensile strength, shear strength, fracture toughness, fracture-matrix coupling ratio, saturation, fracturing-induced fracture threshold, fracturing priority), so as to provide data support for the subsequent training of the intelligent perception model for coal seam structure-physical parameters;

[0060] Construction of an intelligent perception model for coal seam structure-physical parameters: This model takes the CT images of a group of coal seam slices as input and can realize the synchronous estimation of the semantic region segmentation of the coal seam and the corresponding material physical parameters; specifically, the present invention designs two key modules, including a coal seam multi-scale spatial attention feature encoder and a bedding context perception decoder;

[0061] Three-dimensional finite element modeling of the coal seam: Based on the obtained coal seam semantic segmentation and material physical parameters, three-dimensional finite element modeling is carried out; specifically, first, regional stratification and grid division are carried out according to the semantic segmentation information, and then the material physical parameters are accurately mapped to each unit of the finite element model according to the semantic region to obtain a three-dimensional finite element model of the coal seam;

[0062] Agent-assisted dynamic optimization of coal seam hydraulic fracturing parameters: First, construct a multi-objective optimization problem for coal seam hydraulic fracturing parameters; second, construct a surrogate model with embedded coal seam physical information; third, generate training data and test data based on the constructed three-dimensional finite element model of the coal seam to train and test the surrogate model with embedded coal seam physical information to obtain a global surrogate model; finally, combine the global surrogate model with a multi-objective evolutionary algorithm to dynamically solve the multi-objective optimization problem of coal seam hydraulic fracturing parameters and obtain the optimal coal seam hydraulic fracturing parameters.

[0063] The following specifically describes the specific implementation process of the present invention in combination with specific embodiments.

[0064] I. Construction of coal seam semantic segmentation and material calibration data sets

[0065] To achieve coal seam structure semantic segmentation based on coal seam CT images, intelligent perception of material physical parameters, and use them for subsequent construction of coal seam finite element models, it is first necessary to collect coal seam semantic segmentation and material calibration data sets, including the acquisition of initial coal seam slice CT images, as well as semantic category and material physical parameter calibration; specifically, the following steps are included:

[0066] S11. For the coal seam body to be collected, obtain a sequence of slice CT images with spatial continuity; specifically, evenly divide the coal seam body into slices, so each group of CT image data contains slice images collected at equal intervals between layers, that is, , so as to ensure that the CT images can reflect the three-dimensional coal seam block structure; and during the acquisition process, a unified coordinate system calibration is adopted, and the image resolution is ensured to be consistent to support subsequent three-dimensional reconstruction and regional matching;

[0067] S12. According to the characteristics of coal and rock in CT images such as gray-scale distribution, texture morphology, and structural continuity, and combining existing coal seam lithology data and geological structure interpretations, use the LabelMe tool to slice images in the key coal seam geological structures for semantic annotation, including 8 categories: background, main coal seam, interlayer gangue, original fissures, faults / joints, mineralized zones, soft interlayers, and aquifers, and finally obtain semantic segmentation maps of slice images ;

[0068] S13. On the basis of semantic annotation, calibrate the corresponding material physical parameters for each structural category area. The calibrated parameters cover 10 key indicators closely related to stress propagation, fracture induction, seepage characteristics, etc. during the hydraulic fracturing process, including elastic modulus and Poisson's ratio , Permeability , Tensile strength , Shear strength , Fracture toughness , Fracture-matrix coupling ratio , Saturation , Fracture-induced fracture threshold , Fracture priority ; Therefore, for slice images, a total of groups of material physical parameter maps are obtained , and the material physical parameter maps contain the material physical parameters corresponding to each semantic area of the coal seam;

[0069] Repeat the above process, and a total of groups of data jointly constitute the coal seam semantic segmentation and material calibration dataset, and each group of data contains a slice CT image sequence , Semantic segmentation map , and material physical parameter map .

[0070] II. Construction of the intelligent perception model for coal seam structure-physical parameters

[0071] The intelligent perception model for coal seam structure-physical parameters designed by the present invention can intelligently estimate the corresponding coal seam semantic segmentation map and material physical parameter map based on the slice CT image sequence; the model mainly consists of two key modules, including a coal seam multi-scale spatial attention feature encoder and a bedding context perception decoder; the overall model framework is as Figure 2 shown, and the model construction and acquisition process specifically includes the following steps:

[0072] S21, for the th slice CT image of each layer to be segmented , where ; First, two 3*3 convolutional blocks are used to perform preliminary feature extraction on the CT image. Each 3*3 convolutional block consists of a convolutional layer with a convolution kernel of 3*3, a batch normalization layer, and a ReLU activation function, and is used to extract local texture information and basic spatial features; Subsequently, three max-pooling layers are introduced in sequence for multi-scale downsampling to obtain the coal seam structure features at different spatial scales, and the feature outputs from shallow to deep are , corresponding to the front-layer feature, middle-layer feature, and deep-layer feature respectively;

[0073] S22, after completing the preliminary feature extraction, for the multi-scale features Design a multi-scale spatial attention feature encoder for coal seams to achieve deep feature extraction of coal seam structures and the fusion of bedding context information. The encoder consists of a hybrid feature encoding block, an upsampling operation, an average pooling layer, and a spatial attention layer. The specific process is as follows:

[0074] (1)Feature fusion from deep to shallow: After upsampling the deep features , they are fused with the middle-layer features through the second hybrid feature encoding block to obtain the fused feature . Subsequently, continue to upsample and further fuse it with the shallow features through the first hybrid feature encoding block to obtain the shallow-layer fusion result ;

[0075] (2)Feature enhancement and fusion from shallow to deep: After average pooling the shallow-layer fusion result , it is input together with the fused feature into the third hybrid feature encoding block to enhance the medium-scale feature perception ability and obtain the fused feature . Subsequently, further average pool and input it together with the deep features into the fourth hybrid feature encoding block to output the fusion result ;

[0076] (3)Feature integration and spatial attention modeling: Concatenate the fused features , , of three different scales along the channels to construct a multi-scale feature set. Finally, enhance the spatial distribution characteristics of the key structures (including coal seam interfaces and fracture boundaries) in the image through the spatial attention module, and finally output the bedding structure feature map representing the th slice image;

[0077] Among them, the hybrid feature encoding block is used to fuse multi-level features of coal seam images between different scales, enhancing the model's perception ability of complex geological structures. For two groups of features to be fused as input, first, they are concatenated in the channel dimension to form a joint feature representation; subsequently, this joint feature is simultaneously input into three parallel paths to extract spatial and semantic information from different perspectives: Path 1 adopts a residual connection structure with batch normalization, which is used to retain the original feature trends and boundary information, enhancing the stability and transferability of feature fusion; Path 2 first compresses the channels through a 1*1 convolutional block, and then extracts local structural features through a max pooling operation to enhance the model's response ability to key regions such as the main coal seam boundary and parting block; Path 3 performs channel mapping through a 1*1 convolutional block, then introduces a RepVGG layer with deep modeling ability to obtain a stronger semantic representation, and finally further improves the perception accuracy of spatial bedding changes through a 3*3 convolutional layer. Finally, the features extracted from the three paths are concatenated in the channel dimension to form a fused feature representation;

[0078] S23, for each slice CT image , respectively passing through S1 and S2 to obtain the corresponding bedding structure features corresponding to each CT image ; Since the coal seam slices are collected according to the principle of equidistant interlayer sampling, and the coal seam bedding structure has high continuity in physical space, there is significant context semantic correlation between adjacent slices; to make full use of this interlayer continuity, the present invention designs a bedding context-aware decoder to achieve more accurate and consistent image semantic reconstruction by combining the bedding information of adjacent slices;

[0079] Specifically, for the bedding structure features of the th slice , combining the bedding structure features of its two adjacent slices before and after and for channel concatenation to form a context fusion feature (when , select and as its adjacent bedding structure features; when , select and as its adjacent bedding structure features); then, the context fusion feature is input into a 3*3 transposed convolutional block for preliminary decoding upsampling, and further introduces a multi-head attention mechanism to explicitly model and feature enhance the key structures in the bedding context, improving the spatial recognition ability of complex geological structures of the coal seam (including faults, partings, fractures), and obtaining an enhanced feature ;

[0080] Subsequently, decoding is performed through four parallel feature reconstruction paths, each path extracting information from different receptive fields and modeling strategies. Among them, the first path combines deconvolution blocks of 3*1 and 1*3 to enhance the directional perception ability of the bedding transverse and longitudinal boundary structures; the second path uses deconvolution blocks of 3*3 and 1*1 to fuse local details and global contour information; the third path uses a combination of deconvolution blocks of 5*5 and 1*1 to expand the receptive field and capture larger-scale bedding change features; the fourth path effectively retains the original semantic features and strengthens the deep structure information through the residual connection mechanism. Finally, the outputs of the four paths are concatenated in channels to obtain the decoder output features for the bedding structure feature map of the th slice image ;

[0081] S24, finally, through the output layer, is feature compressed and separately mapped to obtain the estimation results of the final semantic segmentation maps of the slice CT images, and the estimation results of the material physical parameter maps ;

[0082] S25, based on the constructed coal seam semantic segmentation and material calibration dataset, the intelligent perception model of coal seam structure-physical parameters is trained. During the training process, the cross-entropy loss function is used to calculate the classification error of the semantic segmentation map, and the mean square error (MSE) loss function is used to calculate the regression error of the material physical parameter map; the Adam optimization algorithm is used for parameter update during model training, and finally, the trained intelligent perception model of coal seam structure-physical parameters is obtained.

[0083] III. 3D finite element modeling of coal seam

[0084] Based on the estimation results of the obtained semantic segmentation map and material physical parameter map, a 3D finite element model that can truly reflect the coal seam structure distribution and material physical properties is constructed to support the numerical simulation and parameter optimization of the subsequent hydraulic fracturing process; specifically, it includes the following steps:

[0085] S31, 3D structure reconstruction and regional stratification modeling: First, based on the estimation results of the semantic segmentation map of the sequential CT slice images 3D structure reconstruction is performed; the 3D semantic structure distribution of the coal seam body is obtained by arranging the equidistant slice images in sequence in a unified coordinate system. For different semantic categories (including main coal seam, parting, fault, fracture), corresponding structural voxel models are constructed, and regional stratification is performed to obtain the stratified semantic structure model;

[0086] S32, Mesh Generation and Structural Discretization: Based on the hierarchical semantic structure model, three-dimensional finite element method is used for structural discretization. For complex geological structure forms (including original fissures, faults / joints, mineralized zones, and weak interlayers), unstructured mesh generation strategy is adopted, and the element density is adaptively adjusted according to the geometric characteristics of different regions. Finally, a discretized finite element model is obtained;

[0087] S33, Mapping of Material Physical Parameters: Map various parameters (such as elastic modulus, permeability, shear strength, etc.) in the material physical parameter map estimation result accurately to the finite element mesh cells according to the semantic regions; this mapping process supports pixel-by-pixel parameter correspondence to ensure that the model can delicately reflect the heterogeneity and spatial variability of the physical properties inside the coal seam, and a three-dimensional finite element model of the coal seam is obtained , representing material physical parameters.

[0088] IV. Dynamic Optimization Process of Coal Seam Hydraulic Fracturing Parameters Based on Surrogate-Assisted

[0089] 1. Construction of Multi-Objective Optimization Problem for Coal Seam Hydraulic Fracturing Parameters

[0090] In the present invention, maximizing the stimulation volume, minimizing the comprehensive cost, and maximizing the fracture conductivity are taken as the optimization objectives, and an engineering parameter in the hydraulic fracturing process is taken as the decision variable to establish a multi-objective optimization problem for coal seam hydraulic fracturing parameters, which is specifically as follows:

[0091] Determine the decision variables: In order to optimize the key parameters in the whole process of coal seam hydraulic fracturing, the present invention selects 9 key parameters as decision variables, including 6 fracturing construction parameters and 3 engineering control parameters. Among them, the fracturing construction parameters include displacement , whose engineering significance is to control the fracture propagation speed; sand ratio , whose engineering significance is to affect the proppant transportation efficiency; fracturing fluid viscosity , whose engineering significance is to affect the sand-carrying capacity and filtration loss. Injection time , whose engineering significance is to control the fracture extension duration; stage volume , used to control the single-stage fracturing scale; pumping rate , whose engineering significance is to affect the construction continuity. The engineering control parameters include perforation cluster spacing , whose engineering significance is to control the fracture initiation position; proppant particle size , whose engineering significance is to balance the conductivity and settlement; temporary plugging agent concentration , whose engineering significance is to affect the multi-cluster extension balance;

[0092] Design the objective function: The objective function is specifically as follows:

[0093] ;

[0094] Among them is the target function of the transformation volume. Minimizing this target function helps to improve the utilization rate of natural fractures in coal seams. is the transformation volume, which is calculated by the three-dimensional finite element model of the coal seam; is the comprehensive cost target function, which covers the cost of sand materials, the cost of fracturing fluid and the cost of temporary plugging agent consumption. Minimizing this target function helps to reduce the cost of hydraulic fracturing in coal seams. ， ， and are the fixed cost coefficient of displacement, the fixed cost coefficient of sand ratio, the fixed cost coefficient of stage volume and the fixed cost coefficient of temporary plugging agent concentration respectively; is the target function of fracture conductivity. Minimizing this target function helps to improve the fracture conductivity after hydraulic fracturing of the coal seam. is the conductivity coefficient, is the fracture permeability, which is calculated by the three-dimensional finite element model of S coal seam;

[0095] 2. Construct a surrogate model embedded with coal seam physical information

[0096] Using the three-dimensional finite element model of the coal seam to calculate and requires a long simulation time. Moreover, since there is no time-dependent relationship and spatial-dependent relationship among the nine decision variables to be optimized in the present invention, traditional time series models and spatial models are difficult to fully learn the potential correlation among the nine decision variables and the contribution of the decision variables to and the predicted value. To reduce the simulation time cost of the finite element model and improve the optimization efficiency, the present invention constructs a surrogate model embedded with coal seam physical information to learn the mapping relationship between the decision variables to and The model structure is as Figure 3 shown. The surrogate model embedded with coal seam physical information includes an encoder, a large-scale physical information extraction layer, a small-scale physical information extraction layer, a physical information aggregation layer and a fully connected layer. The data= ; ; ; ; ; ; ​; ; Obtain the surrogate model embedded with the physical information of the input coal seam and predicted values. The specific steps are as follows:

[0097] 1) Input data into the encoder to extract the primary features of each decision variable data, and obtain the decision variable primary feature matrix ; The encoder consists of a convolutional layer and a RELU activation function layer. The convolutional layer is used to extract the local features of each decision variable data, and the RELU activation function layer is used to perform non-linear transformation on the local features of each decision variable data to enhance the generalization ability of the model;

[0098] 2) Input data into the large-scale physical information extraction layer to extract the large-scale correlation features of each decision variable data, and obtain the large-scale correlation feature matrix , The large-scale physical information extraction layer uses a graph attention neural network with a convolutional kernel scale of to extract the large-scale features of each decision variable data in data, and calculate the correlation coefficients for the large-scale features of each decision variable to obtain ;

[0099] 3) Input data into the small-scale physical information extraction layer to extract the small-scale correlation features of each decision variable data, and obtain the small-scale correlation feature matrix , The large-scale physical information extraction layer uses a graph attention neural network with a convolutional kernel scale of to extract the small-scale features of each decision variable data in data, and calculate the correlation coefficients for the small-scale features of each decision variable to obtain ;

[0100] 4) Add and , and then input the result together with into the physical information aggregation layer to aggregate the graph features and local features to obtain the aggregated feature ; The physical information aggregation layer uses a graph aggregation neural network to sample and aggregate the features of each decision variable in F and the features of each decision variable in the feature map obtained by adding and respectively, to obtain the aggregated feature representing the implicit relationship of the decision variable data;

[0101] 5) Input into the fully connected layer to perform dimensionality reduction and mapping on the features to obtain and predicted values;

[0102] 3. Initialize the surrogate model

[0103] Use the Latin hypercube sampling method to sample respectively for 、 、 、 、 、 、 、 and within their value ranges for times of sampling to obtain groups of engineering parameter data; Take the groups of engineering parameter data as the parameters of coal seam hydraulic fracturing simulation in sequence, and perform time-domain simulation on the established three-dimensional finite element model of the coal seam to obtain the reformed volume and fracture permeability corresponding to the groups of engineering parameter data. After merging the groups of engineering parameter data and their corresponding reformed volume and fracture permeability, divide them into a training set and a test set according to a ratio of 8:2. Use the training set and the test set to train and test the surrogate model embedded with coal seam physical information respectively, and save the optimal model structure to obtain the global surrogate model;

[0104] 4. Dynamically solve the multi-objective optimization problem of coal seam hydraulic fracturing parameters

[0105] The present invention combines the obtained global surrogate model with a multi-objective evolutionary algorithm to dynamically solve the multi-objective optimization problem of coal seam hydraulic fracturing parameters. The specific steps are as follows:

[0106] 1) First, use the Latin hypercube sampling to sample respectively for 、 、 、 、 、 、 、 and sample within their value ranges for times to obtain Group decision variables; secondly, each group of decision variables is used as the parameters of the coal seam hydraulic fracturing simulation in turn, and the time-domain simulation is performed on the three-dimensional finite element model of the coal seam to calculate the transformation volume and fracture permeability corresponding to each group of decision variables; thirdly, each group of decision variables and their corresponding transformation volume and fracture permeability are brought into the objective function to calculate the objective value of each group of decision variables; finally, the decision variables and the objective values are combined to obtain the population , where the th population individual , where is the th objective value vector of the population individual ;

[0107] 2) Initialize the knowledge base and store in the knowledge base;

[0108] 3) Initialize the evaluation times and set the maximum evaluation times;

[0109] 4) Perform binary tournament selection on to obtain the paired population ;

[0110] 5) Perform crossover and mutation operations on to obtain the offspring population ;

[0111] 6) Use the global surrogate model obtained by S5-3 to predict the transformation volume and fracture permeability of all population individuals in , and bring the decision variables of the population individuals and their corresponding transformation volume and fracture permeability into the objective function described in S5-1 to calculate the objective values of the population individuals;

[0112] 7) Perform non-dominated sorting and crowding distance calculation on to screen out the population located on the first Pareto front ; Take the decision variables of each population individual in as the parameters of the coal seam hydraulic fracturing simulation in turn, and perform time-domain simulation on the three-dimensional finite element model of the coal seam to calculate the transformation volume and fracture permeability corresponding to each group of decision variables; again, bring the decision variables of the population individuals and their corresponding transformation volume and fracture permeability into the objective function to calculate the objective values of the population individuals;

[0113] 8) Update the evaluation times and add the evaluation times to the scale of ;

[0114] 9) Store in the knowledge base; when the number of population individuals in the knowledge base is greater than When it is time, first, perform non-dominated sorting and cosine similarity calculation on the population in the knowledge base. Secondly, use the cosine similarity to pair the population individuals. Finally, remove the population individuals with a larger non-dominated sorting from the knowledge base for each pair of population individuals;

[0115] 10) Combine and to obtain a combined population, perform non-dominated sorting and crowding distance calculation on the combined population, and select the population individuals with a sorting of and those less than it to form the next-generation population for update ;

[0116] 11) Judge whether the number of evaluations is greater than the maximum number of evaluations. If it is greater, output , otherwise repeat steps 4)-11);

[0117] The output in step 11) is the optimal Pareto solution set for the multi-objective optimization problem of coal seam hydraulic fracturing parameters. Select the decision variables of the population individual with the smallest L2 norm between the target value and the origin in as the final optimization parameters.

[0118] V. Simulation Experiment

[0119] 1. Verify the performance of the intelligent perception model for coal seam structure-physical parameters

[0120] To verify the effect of the intelligent perception model for coal seam structure-physical parameters in coal seam structure semantic segmentation, the present invention selects multiple groups of representative coal seam slice CT images as inputs and outputs their corresponding structure semantic segmentation results. In addition, since the material physical parameter map is a further mapping based on semantic region recognition, the experiment focuses on reflecting the model's ability to recognize coal seam structures through the semantic segmentation results. The experimental results are as Figure 4 shown, where the pink is the main coal seam, the blue-green is the interlayer parting, the yellow is the soft interlayer, and the black is the original fracture;

[0121] Since the multi-scale spatial attention feature encoder can capture the geological structure differences at different scales in the coal seam image and strengthen the representation of the features of key structures (such as the main coal seam, parting, and weak interlayers) through the spatial attention mechanism, the contour of the main coal seam (pink) area in the segmentation result is clear and the shape is coherent. Moreover, the parting (blue-green) and weak interlayers (yellow) can be accurately distinguished in the structural transition area. This result verifies the advantage of multi-scale modeling in fine-grained structure division. In addition, due to the introduction of context bedding information in the bedding context-aware decoder, the model can still maintain accurate discrimination of the structural boundaries under complex textures, thus ensuring that the model still has excellent detection effects on complex textures such as original fractures (black). Therefore, the experiment shows that the intelligent perception model of coal seam structure-physical parameters exhibits good segmentation accuracy and structure discrimination ability in complex coal seam images with multiple structures;

[0122] 2. Verify the performance of the proxy model with coal seam physical information embedding proposed in the present invention combined with the multi-objective evolutionary algorithm to solve the multi-objective optimization problem of coal seam hydraulic fracturing parameters proposed in the present invention

[0123] Establish a multi-objective optimization problem of coal seam hydraulic fracturing parameters based on the data of a coalfield in the western region of Shandong Province. Use the method of the present invention (PINN-MOEAS), NSGA-II based on the Kriging model (Kriging-NSGA-II), and MOPSO based on the Kriging model (Kriging-MOPSO) to solve the multi-objective optimization problem of coal seam hydraulic fracturing parameters. The hypervolume convergence curve is as Figure 5 shown.

[0124] It can be seen from Figure 5 that when the number of function evaluations (FEs) reaches 600, the hypervolume value (HV) of the method of the present invention is greater than that of NSGA-II based on the Kriging model and MOPSO based on the Kriging model, which proves that the method of the present invention can improve the convergence speed of the population and the efficiency of solving the multi-objective optimization problem of coal seam hydraulic fracturing parameters, and proves that the proxy model with coal seam physical information embedding adopted in the present invention has high prediction accuracy and can accurately guide the population to search. The two-dimensional visualization and three-dimensional visualization of the best population obtained by the method of the present invention on the three optimization objectives are as Figure 6 shown. Among them, Figure (a) is the two-dimensional visualization scatter plot of population objective value 1 ( ), and population objective value 2 ( ), and Figure (b) is the two-dimensional visualization scatter plot of population objective value 1 and population objective value 3 ( The two-dimensional visualization scatter plot of ( ) and the three-dimensional visualization scatter plot of population objective value 1, population objective value 2, and population objective value 3 are shown in Figure (c). It can be seen that the solution sets obtained by the method of the present invention are evenly distributed in space both in the two-dimensional scatter plot and the three-dimensional scatter plot, proving that the method of the present invention can effectively maintain the diversity of the solution sets.

[0125] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0126] Although the specific implementation manners of the present invention are described above, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning, characterized in that: The following steps are involved: S1, collect CT images of coal seam slices with spatial continuity; S2, input the CT image into the constructed and trained coal seam structure-physical parameter intelligent perception model, which includes a coal seam multi-scale spatial attention feature encoder and a bedding context perception decoder, which is used to output the coal seam semantic segmentation results and the corresponding material physical parameters; S3, performing regional stratification and meshing according to the semantic segmentation results, and then mapping the corresponding material physical parameters to each unit of the finite element model according to the semantic region to obtain a three-dimensional finite element model of the coal seam; S4, taking maximizing the transformation volume, minimizing the comprehensive cost and maximizing the fracture conductivity as the optimization objectives, and taking the coal seam hydraulic fracturing parameters to be solved as the decision variables, a multi-objective optimization problem of coal seam hydraulic fracturing parameters is established; the multi-objective optimization problem is solved based on the agent-assisted multi-objective evolutionary algorithm to obtain the optimal coal seam hydraulic fracturing parameters; The agent-assisted multi-objective evolutionary algorithm generates training data and test data based on the constructed three-dimensional finite element model of the coal seam to train and test the agent model embedded with the physical information of the coal seam to obtain a global agent model, and combines the global agent model with the multi-objective evolutionary algorithm to dynamically solve the multi-objective optimization problem.

2. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning as claimed in claim 1, characterized in that: Based on the model training task, the coal seam semantic segmentation and material calibration datasets are collected and constructed. The specific process is as follows: First, for the coal seam to be collected, a slice CT image sequence with spatial continuity is obtained: the coal seam is evenly divided into slices, each set of CT image data contains The slice images collected at equidistant intervals are ,The acquisition process adopts a unified coordinate system calibration; Then The key coal seam geological structures in the slice images are semantically annotated, including 8 categories: background, main coal seam, interlayer gangue, original cracks, faults / joints, mineralized zones, weak interlayers, and aquifers. Semantic segmentation map of slice images ; Finally, based on the semantic annotation, the corresponding material physical parameters of each structural category area are calibrated. The calibrated parameters include elastic modulus , Poisson's ratio , Permeability , tensile strength , Shear Strength , fracture toughness , fracture-matrix coupling ratio , Saturation , Fracturing-induced fracture threshold , Fracturing Priority ;against slice images, a total of Group material physical parameter diagram ,The material physical parameter map contains the material physical parameters corresponding to each coal seam semantic area.

3. The method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning according to claim 1, characterized in that: The specific processing process of the layer structure-physical parameter intelligent perception model is as follows: S21, for Each slice of CT image to be segmented ,in ; First, two 3*3 convolution blocks are used to perform preliminary feature extraction on the CT image. Each 3*3 convolution block consists of a convolution layer with a convolution kernel of 3*3, a batch normalization layer, and a ReLU activation function to extract local texture information and basic spatial features; Subsequently, three maximum pooling layers are introduced in turn for multi-scale downsampling to obtain the coal seam structural characteristics at different spatial scales, and the feature outputs from shallow to deep are: , corresponding to the front-layer features, middle-layer features and deep-layer features respectively; S22, multi-scale features Input coal seam multi-scale spatial attention feature encoder to realize deep feature extraction of coal seam structure and fusion of bedding context information: For deep features After upsampling, compared with the middle-level features The fusion feature is obtained by fusing through the second mixed feature encoding block ; then continue to Upsample and combine with shallow features Further fusion is performed through the first hybrid feature encoding block to obtain a shallow fusion result ; Shallow fusion results After average pooling, combined with the fusion feature The third hybrid feature encoding block is input together to enhance the perception of mid-scale features and obtain fused features ; then further Perform average pooling and combine with deep features Input to the fourth hybrid feature encoding block and output the fusion result ; The fusion features of three different scales , , Perform channel splicing and construct a multi-scale feature set; Finally, the spatial attention module is used to enhance the spatial distribution characteristics of the key structures in the image. The final output represents Layer structure feature map of the slice image ; S23, based on the bedding context-aware decoder, performs image semantic reconstruction by combining the bedding information of adjacent slices: For the Bedding structure characteristics of each slice , combined with the bedding structure characteristics of the two adjacent slices and Perform channel splicing to form context fusion features ; Afterwards, the context fusion feature It is input into a 3*3 deconvolution block for preliminary decoding and upsampling, and a multi-head attention mechanism is introduced to explicitly model and enhance the key structures in the layered context to obtain enhanced features. ; Then, After decoding through four parallel feature reconstruction channels, each channel extracts information from different receptive fields and modeling strategies, and the outputs of the four channels are spliced ​​to obtain the Slice image layer structure feature map The decoder output features ; S24, and finally through the output layer Perform feature compression and map them separately The estimated result of the final semantic segmentation map of the slice CT image , and the estimated results of the material physical parameter map .

4. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning as claimed in claim 3, characterized in that: The first mixed feature coding block, the second mixed feature coding block, the third mixed feature coding block, and the fourth mixed feature coding block are specifically: For the two sets of input features to be fused, they are first concatenated in the channel dimension to form a joint feature representation; then, the joint feature is simultaneously input into three parallel pathways to extract spatial and semantic information from different angles: Path 1 uses a residual connection structure with batch normalization to preserve the original feature trend and boundary information; Path 2 first compresses the channel through a 1*1 convolution block, and then extracts local structural features through a maximum pooling operation to enhance the model's responsiveness to the key areas of the main coal seam boundary and interbedded gangue blocks; Pathway 3 performs channel mapping through a 1*1 convolution block, then introduces the RepVGG layer with deep modeling capabilities to obtain stronger semantic representation, and finally uses a 3*3 convolution layer to further improve the perception accuracy of spatial layer changes; finally, the features extracted from the three paths are channel-joined to form a fused feature representation.

5. The method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning according to claim 3, characterized in that: Said Decoding is performed through four parallel feature reconstruction paths, specifically: Among them, the first pathway uses a combination of 3*1 and 1*3 deconvolution blocks to enhance the directional perception of the horizontal and vertical boundary structures of the bedding; the second pathway uses 3*3 and 1*1 deconvolution blocks to fuse local details with global contour information; the third pathway uses a combination of 5*5 and 1*1 deconvolution blocks to expand the receptive field and capture larger-scale bedding change characteristics; the fourth pathway effectively retains the original semantic features and strengthens the deep structural information through the residual connection mechanism.

6. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning as claimed in claim 1, characterized in that: The specific process of S3 is as follows: S31, 3D structure reconstruction and regional hierarchical modeling: First, the semantic segmentation map of the serial CT slice images is used to estimate the results. Reconstruct the three-dimensional structure; obtain the three-dimensional semantic structure distribution of the coal seam by arranging the equidistant slice images in sequence in a unified coordinate system, construct corresponding structural voxel models for different semantic categories, and perform regional stratification to obtain a hierarchical semantic structure model; S32, Meshing and structural discretization: Based on the hierarchical semantic structural model, the three-dimensional finite element method is used for structural discretization, and an unstructured meshing strategy is adopted for complex geological structural morphology. The unit density is adaptively adjusted in combination with the geometric characteristics of different regions, and finally a discretized finite element model is obtained; S33, Material physical parameter mapping: Mapping the material physical parameter estimation results Various parameters in are accurately mapped to the finite element mesh units according to semantic regions; The mapping process supports pixel-by-pixel parameter correspondence, so that the model can carefully reflect the heterogeneity and spatial variability of the physical property distribution inside the coal seam, and obtain a three-dimensional finite element model of the coal seam. , Represents the physical parameters of the material.

7. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning as claimed in claim 1, characterized in that: The multi-objective optimization problem of coal seam hydraulic fracturing parameters established in S4 is specifically: Determine the decision variables: 9 key parameters were selected as decision variables, including 6 fracturing construction parameters and 3 engineering control parameters; the fracturing construction parameters include displacement , its engineering significance is to control the crack expansion speed; sand ratio , its engineering significance is to affect the proppant delivery efficiency; fracturing fluid viscosity , its engineering significance is to affect the sand carrying capacity and filtration loss; Injection time , its engineering significance is to control the extension time of cracks; Stage Volume , used to control the scale of single-stage fracturing; pumping rate , its engineering significance is to affect the construction continuity; engineering control parameters include perforation cluster spacing , its engineering significance is to control the crack initiation position; proppant particle size , its engineering significance is to balance the flow conductivity and sedimentation; the concentration of temporary plugging agent , its engineering significance is the balance of image multi-cluster expansion; Design objective function: ； in To transform the volume objective function, the objective function is minimized to improve the utilization rate of natural fractures in coal seams. is the transformation volume, which is calculated from the three-dimensional finite element model of the coal seam; is a comprehensive cost objective function, which covers the cost of sand material, fracturing fluid and temporary plugging agent consumption. Minimizing this objective function is used to reduce the cost of coal seam hydraulic fracturing. ， ， and They are fixed cost coefficient of displacement, fixed cost coefficient of sand ratio, fixed cost coefficient of stage volume and fixed cost coefficient of temporary plugging agent concentration; is the fracture conductivity objective function, and minimizing this objective function is used to improve the fracture conductivity of the coal seam after hydraulic fracturing. is the conductivity coefficient, is the fracture permeability, which is calculated by the three-dimensional finite element model of the coal seam.

8. A method for dynamic optimization of coal seam hydraulic fracturing parameters based on deep learning as claimed in claim 7, characterized in that: The global proxy model includes an encoder, a large-scale physical information extraction layer, a small-scale physical information extraction layer, a physical information aggregation layer and a fully connected layer; data Enter the global proxy model to obtain and Prediction value, the specific steps are as follows: Input data into the encoder to extract the primary features of each decision variable data and obtain the primary feature matrix of the decision variable ; The encoder is composed of a convolution layer and a RELU activation function layer, the convolution layer is used to extract the local features of each decision variable data, and the RELU activation function layer is used to perform nonlinear transformation on the local features of each decision variable data; Input data into the large-scale physical information extraction layer to extract the large-scale correlation features of each decision variable data and obtain the large-scale correlation feature matrix The large-scale physical information extraction layer uses a convolution kernel scale of The graph attention neural network is used to extract the large-scale features of each decision variable in the data and calculate the correlation coefficient of the large-scale features of each decision variable to obtain ; Input data into the small-scale physical information extraction layer to extract the small-scale correlation features of each decision variable data and obtain the small-scale correlation feature matrix The large-scale physical information extraction layer uses a convolution kernel scale of The graph attention neural network is used to extract the small-scale features of each decision variable in the data and calculate the correlation coefficient of the small-scale features of each decision variable to obtain ; Will and After adding Input the physical information aggregation layer together to aggregate the graph features and local features to obtain aggregated features The physical information aggregation layer uses a graph aggregation neural network to combine the characteristics of each decision variable in F with and The features of each decision variable in the added feature graph are sampled and aggregated respectively to obtain the aggregated features that characterize the implicit relationship between the decision variable data. ; Will Input the fully connected layer to reduce the dimension and map the features and Predicted value.

Images