A shale fracture modeling platform based on cloud computing

Abstract

The present invention relates to the technical field of cloud computing applications, and discloses a shale fracture modeling platform based on cloud computing. The platform includes a multi-source data acquisition module, a heterogeneous data fusion module, a fracture dynamic modeling module, a parameter optimization module, and a cloud verification module. The multi-source data acquisition module collects various data and uploads it to the cloud; the heterogeneous data fusion module processes data heterogeneity to generate a fusion feature matrix; the fracture dynamic modeling module constructs a three-dimensional dynamic model; the parameter optimization module optimizes physical properties parameters; the cloud verification module verifies the model in multiple dimensions and manages versions. By leveraging the advantages of cloud computing, the platform solves the problems of traditional modeling in data acquisition, fusion, modeling algorithms, parameter optimization, and model verification, can efficiently process multi-source data, accurately construct a fracture model, improve the accuracy and reliability of the model, realize comprehensive verification and effective management of the model, and provide strong support for the efficient exploitation of shale gas.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing applications, and particularly to a shale fracture modeling platform based on cloud computing. Background Art

[0002] With the continuous growth of global energy demand, shale gas, as an important unconventional natural gas resource, has attracted much attention for its efficient development. The exploitation effect of shale gas depends to a large extent on the accurate understanding and modeling analysis of shale fractures, which is directly related to reserve assessment, exploitation plan design, and recovery rate improvement. However, there are still many challenges in the field of shale fracture modeling.

[0003] The data sources required for shale fracture modeling are extensive, covering seismic wave data, core physical property data, drilling trajectory data, microseismic monitoring data, historical fracturing construction data, etc. Traditional data acquisition methods are not only inefficient, but also due to the limitations of data acquisition equipment and the complexity of the acquisition environment, the obtained data often has problems such as low accuracy, data missing or errors. For example, in remote shale gas exploitation areas, the signals of sensors are easily interfered, resulting in noise in the collected seismic wave data and affecting data quality. At the same time, the formats and structures of these multi-source data are extremely different, bringing great difficulties to the centralized storage and unified management of data. Different types of data are stored in different systems or media respectively, and the integration and invocation of data are extremely inconvenient, seriously restricting the development of subsequent modeling work.

[0004] Due to the different acquisition times, locations, and methods of various types of data, there are obvious spatio-temporal differences and data characteristic differences between them, that is, heterogeneity. In the time dimension, the acquisition frequencies of seismic wave data and microseismic monitoring data are different, resulting in difficulty in aligning time series; in the space dimension, the sampling point distributions of drilling trajectory data and core physical property data are uneven and cannot be directly fused for analysis. In addition, the feature dimensions of the data are also different. For example, core physical property data contains complex information such as various mineral components and mechanical parameters, with a relatively high dimension, while some monitoring data has a relatively low dimension. Existing data fusion technologies are difficult to effectively solve these problems and cannot fully explore the potential connections between multi-source data, making the fused data unable to accurately reflect the true characteristics of the shale formation and affecting the accuracy of modeling.

[0005] Traditional shale fracture modeling algorithms have many deficiencies when dealing with complex geological conditions. Methods based on empirical formulas or simple physical models oversimplify geological processes and cannot accurately describe the complex topological structure and dynamic evolution process of shale fractures. For example, these methods usually assume that fractures expand uniformly in homogeneous media, ignoring the influence of formation heterogeneity on fracture propagation. Numerical simulation-based methods, although able to consider more geological factors to a certain extent, have high computational costs and strong dependence on model parameters. When the model parameters are inaccurate, the reliability of the simulation results will be greatly reduced. In recent years, machine learning algorithms have been applied in shale fracture modeling, but still face problems such as poor model generalization ability and difficulty in processing large-scale data. For example, deep learning models are prone to overfitting during training and cannot accurately predict fracture morphology and distribution when facing new geological data.

[0006] Shale fracture models contain numerous physical property parameters, such as fracture permeability, porosity, and stress field distribution parameters. The accuracy of these parameters directly affects the accuracy of the model. However, due to the uncertainty of geological conditions and data limitations, it is very difficult to obtain accurate model parameters. Traditional parameter optimization methods, such as trial-and-error methods or simple optimization algorithms, are inefficient and easily fall into local optimal solutions. These methods cannot fully utilize the information in multi-source data and are difficult to achieve collaborative optimization of multiple parameters, resulting in a large deviation between the model and the actual geological situation.

[0007] In terms of model verification, most existing verification methods only test some features of the model and lack multi-dimensional and systematic verification. For example, some methods only focus on whether the geometric morphology of fractures conforms to actual observations, while ignoring the rationality of physical property parameters; some other methods consider some physical property parameters, but lack quantitative analysis of model uncertainty. In addition, due to the lack of an effective data sharing mechanism and a unified verification standard, it is difficult to compare and evaluate the model verification results in different research or mining projects, and the reliability and universality of the model cannot be guaranteed. Summary of the Invention

[0008] The purpose of the present invention is to provide a cloud computing-based shale fracture modeling platform to solve the problems proposed in the above background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A cloud computing-based shale fracture modeling platform, the platform includes a multi-source data acquisition module, a heterogeneous data fusion module, a fracture dynamic modeling module, a parameter optimization module, and a cloud verification module;

[0010] The multi-source data acquisition module is used to obtain the original multi-source data set required for shale fracture modeling. Specifically, by deploying a distributed sensor network and geological exploration equipment, it collects seismic wave data, core physical property data, drilling trajectory data, microseismic monitoring data, and historical fracturing construction data of the shale formation, and uploads the original multi-source data set to the cloud storage cluster;

[0011] The heterogeneous data fusion module is used to perform cross-domain fusion and feature enhancement on the original multi-source data. Specifically, it eliminates the data acquisition time delay difference through a spatio-temporal alignment algorithm, adopts a multi-scale decomposition method based on wavelet transform to denoise the seismic wave data, and combines a principal component analysis algorithm to reduce the dimensionality of the core physical property data to generate a fusion feature matrix, and transmits the fusion feature matrix to the fracture dynamic modeling module;

[0012] The fracture dynamic modeling module is used to construct a three-dimensional dynamic model of shale fractures. Specifically, it uses a fracture topology generation algorithm based on a graph convolutional network to perform spatial structure modeling on the fusion feature matrix to generate an initial fracture network diagram, and introduces a generative adversarial network to dynamically optimize the fracture distribution to obtain a shale fracture dynamic model, where the discriminator of the generative adversarial network uses an adaptive attention mechanism to optimize the weight distribution;

[0013] The parameter optimization module is used to iteratively optimize the physical property parameters of the fracture model. Specifically, it uses a hybrid particle swarm optimization algorithm and combines a Bayesian probability model to perform multi-objective optimization on the fracture permeability, porosity, and stress field distribution parameters to generate an optimized parameter set, and feeds the optimized parameter set back to the fracture dynamic modeling module for model update;

[0014] The cloud verification module is used to perform multi-dimensional verification on the optimized fracture model. Specifically, by constructing a reconstruction error evaluation model based on a variational autoencoder, it performs consistency verification on the fracture geometry and physical property parameters, and at the same time combines the Monte Carlo simulation method to quantitatively analyze the model uncertainty, outputs the verification result and stores it in the cloud database.

[0015] Preferably, in the multi-source data acquisition module, the seismic wave data includes the P-wave to S-wave velocity ratio, reflection coefficient matrix, and dispersion curve; the core physical property data includes the mineral composition percentage, Young's modulus, and fracture toughness value; the historical fracturing construction data includes the pump injection pressure curve, proppant concentration distribution, and fracturing fluid viscosity time series data.

[0016] Preferably, in the heterogeneous data fusion module, the spatio-temporal alignment algorithm specifically includes the following steps: performing time series matching on the drilling trajectory data and the microseismic monitoring data based on the dynamic time warping algorithm, using the Kriging interpolation method to complement the spatially missing data, and extracting the frequency domain characteristics of the seismic wave data through Fourier transform to construct a multi-source data tensor in a spatio-temporal unified coordinate system.

[0017] Preferably, in the crack dynamic modeling module, the crack topology generation algorithm of the graph convolutional network specifically includes: discretizing the shale formation into a node graph structure, where the node attributes include lithological characteristics and stress states, and the edge weights are calculated from the distance between nodes and the physical property differences; using multi-layer graph convolutional operators to extract local crack connectivity characteristics, and fusing multi-scale topological information through a skip connection mechanism to generate a weighted crack network graph.

[0018] Preferably, the dynamic optimization process of the generative adversarial network includes: the generator uses a sequence generation model with gated recurrent units to predict the crack propagation path according to the node states of the crack network graph; the discriminator uses a multi-head self-attention mechanism to calculate the rationality score of the crack distribution, and iteratively optimizes the generator parameters through an adversarial loss function.

[0019] Preferably, in the hybrid particle swarm optimization algorithm, the particle position update formula introduces a random perturbation term of the simulated annealing mechanism, specifically:

[0020] where, represents the velocity of particle at the -th iteration; represents the velocity of particle at the -th iteration; represents the inertia weight, which is used to balance the global search and local search capabilities of the algorithm; , represent the learning factors, which are used to adjust the step sizes for the particle to move towards the individual optimal position and the global optimal position; , represent two random numbers between [0, 1]; represents the historical optimal position of particle ; represents the position of particle at the -th iteration; represents the global optimal position; represents the annealing coefficient, which is used to control the amplitude of the random perturbation in the simulated annealing mechanism; represents the cooling rate, which determines the rate at which the random perturbation weakens as the number of iterations increases; represents the number of iterations; A random vector that follows a Gaussian distribution, which is used to enhance the global search ability of the algorithm.

[0021] Preferably, the Bayesian probability model constructs the joint probability distribution of fracture parameters by using the non-parametric kernel density estimation method. Specifically, it smooths the parameter correlation relationship in the historical fracturing data through the kernel function to generate a conditional probability table, and uses the Markov chain Monte Carlo sampling method to obtain the parameter optimization direction.

[0022] Preferably, in the reconstruction error evaluation model of the variational autoencoder, the encoder uses a deep residual network to extract the hidden variables of the fracture morphology, and the decoder uses a deconvolution network to reconstruct the three-dimensional fracture geometry. The loss function includes the weighted sum of the KL divergence term and the mean square error term, which is used to quantify the model reconstruction accuracy.

[0023] Preferably, in the Monte Carlo simulation method, a parameter space sample set is generated by Latin hypercube sampling, and the mapping relationship between the fracture propagation probability and the stress field perturbation is constructed based on the response surface model, and the confidence interval and sensitivity index of the model output are calculated.

[0024] Preferably, the cloud verification module further includes a model version management function. Specifically, a unique identifier is generated for the fracture model after each optimization through the hash algorithm, and the parameter optimization path and verification result are recorded in the blockchain evidence storage system.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] With the help of the distributed sensor network and geological exploration equipment, the multi-source data acquisition module can comprehensively collect various types of key data, greatly improving the efficiency and accuracy of data acquisition. Compared with the traditional acquisition method, this module can obtain seismic wave data, core physical property data, etc. in real time, reducing data loss and errors. For example, in the shale gas extraction area with complex terrain, the distributed sensor network can collect data more stably, reducing the influence of environmental interference. After collection, the original multi-source data set is uploaded to the cloud storage cluster, and the powerful storage capacity of cloud computing is used to realize the centralized management and convenient call of data, solving the problems of scattered data and difficult integration in the traditional storage method, and providing a reliable data basis for the subsequent modeling work.

[0027] The heterogeneous data fusion module effectively solves the problem of heterogeneity of multi-source data through technologies such as spatio-temporal alignment algorithm, wavelet transform for noise reduction, and principal component analysis for feature dimensionality reduction. The spatio-temporal alignment algorithm can accurately eliminate the time delay difference in data acquisition. For example, it matches the time series of drilling trajectories and microseismic monitoring data based on the dynamic time warping algorithm, and uses Kriging interpolation to complement the missing spatial data to ensure the spatio-temporal consistency of the data. The wavelet transform reduces noise in seismic wave data, improves data quality, and highlights effective signals. The principal component analysis reduces the dimensionality of core physical property data, reduces data redundancy while retaining key information, and improves the efficiency of subsequent processing. The generated fusion feature matrix more accurately reflects the true characteristics of the shale formation, laying a foundation for building a high-precision fracture model.

[0028] The fracture dynamic modeling module uses a fracture topology generation algorithm based on graph convolutional network and a generative adversarial network with an adaptive attention mechanism, achieving a major breakthrough in shale fracture modeling. The graph convolutional network discretizes the shale formation into a node graph structure, comprehensively considers lithological characteristics, stress states, and relationships between nodes, effectively extracts local fracture connectivity features, and fuses multi-scale topological information through skip connections to generate a more realistic initial fracture network diagram. During the dynamic optimization process of the generative adversarial network, the generator uses gated recurrent units to predict the fracture propagation path, and the discriminator accurately calculates the rationality score of the fracture distribution through a multi-head self-attention mechanism. The two are trained against each other to make the fracture distribution more reasonable. The constructed three-dimensional dynamic model can more accurately simulate the complex structure and dynamic evolution process of shale fractures, providing a more accurate geological model reference for shale gas exploitation.

[0029] The parameter optimization module combines the hybrid particle swarm optimization algorithm and the Bayesian probability model to achieve efficient iterative optimization of the physical property parameters of the fracture model. The hybrid particle swarm optimization algorithm introduces a random perturbation term of the simulated annealing mechanism to avoid the algorithm falling into local optimal solutions, improves the global search ability, and can explore the parameter space more comprehensively. For example, when optimizing parameters such as fracture permeability, porosity, and stress field distribution, the algorithm can dynamically adjust the search strategy according to the actual situation. The Bayesian probability model constructs a joint probability distribution of fracture parameters through non-parametric kernel density estimation, and uses Markov chain Monte Carlo sampling to obtain the optimization direction, making full use of the information in historical fracturing data, so that the optimized parameters are more in line with the actual geological conditions, significantly improving the accuracy and reliability of the model.

[0030] The cloud verification module conducts multi-dimensional verification on the optimized fracture model by constructing a reconstruction error evaluation model based on variational autoencoders and a Monte Carlo simulation method. The encoder of the variational autoencoder uses a deep residual network to extract latent variables of fracture morphology, and the decoder reconstructs the three-dimensional fracture geometry through a deconvolution network. The loss function comprehensively considers the KL divergence term and the mean square error term to effectively quantify the model reconstruction accuracy and verify the consistency of fracture geometry and physical properties. The Monte Carlo simulation method generates a parameter space sample set through Latin hypercube sampling, constructs a mapping relationship based on the response surface model, calculates the confidence interval and sensitivity index of the model output, and conducts a quantitative analysis of model uncertainty. In addition, the model version management function generates a unique identifier through a hash algorithm, records the parameter optimization path and verification results in the blockchain evidence storage system, realizes the effective management of model versions and data traceability, ensures the stability and traceability of model quality, and provides a reliable basis for the formulation of shale gas exploitation plans.

[0031] The entire platform is built based on a cloud computing architecture, making full use of the powerful computing power and flexible scalability of cloud computing. When processing large-scale multi-source data and complex modeling algorithms, cloud computing can quickly complete computing tasks, greatly shortening the modeling time. For example, when conducting large-scale Monte Carlo simulations, the cloud computing platform can perform parallel computing, significantly improving the computing efficiency. At the same time, the scalability of cloud computing enables the platform to easily handle the increase in data volume and computing tasks, reducing the hardware maintenance cost, improving the stability and reliability of the system, and providing an economical and efficient modeling solution for shale gas exploitation enterprises. Brief Description of the Drawings

[0032] Figure 1 It is the working principle diagram of the shale fracture modeling platform described in the present invention;

[0033] Figure 2 It is the flow chart of shale fracture dynamic modeling based on graph convolutional networks and generative adversarial networks;

[0034] Figure 3 It is the flow chart of shale fracture model parameter optimization based on the hybrid particle swarm optimization algorithm. Detailed Embodiments

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0036] Please refer to Figures 1-3, the present invention provides a technical solution: a shale fracture modeling platform based on cloud computing, the platform includes:

[0037] A multi-source data acquisition module, responsible for obtaining the original multi-source data set required for shale fracture modeling. By deploying a distributed sensor network and geological exploration equipment, seismic wave data, core physical property data, drilling trajectory data, microseismic monitoring data, and historical fracturing construction data of the shale formation are collected. After these data are collected, they will be uploaded to the cloud storage cluster to provide a data basis for subsequent modeling work.

[0038] A heterogeneous data fusion module, which performs cross-domain fusion and feature enhancement on the original multi-source data. First, the time-space alignment algorithm is used to eliminate the time delay difference in data acquisition to ensure the consistency of data in time and space; then, a multi-scale decomposition method based on wavelet transform is used to denoise the seismic wave data to remove noise interference in the data; at the same time, the principal component analysis algorithm is combined to reduce the dimension of the core physical property data, reduce the data dimension, improve the data processing efficiency, and finally generate a fusion feature matrix, which is transmitted to the fracture dynamic modeling module.

[0039] A fracture dynamic modeling module, which constructs a three-dimensional dynamic model of shale fractures based on the fusion feature matrix. A fracture topology generation algorithm based on graph convolutional network is used to perform spatial structure modeling on the fusion feature matrix to generate an initial fracture network diagram; a generative adversarial network is introduced to dynamically optimize the fracture distribution, and the discriminator of the generative adversarial network uses an adaptive attention mechanism to optimize the weight allocation, so as to obtain a shale fracture dynamic model.

[0040] A parameter optimization module, which iteratively optimizes the physical property parameters of the fracture model. The hybrid particle swarm optimization algorithm is used to perform multi-objective optimization on the fracture permeability, porosity, and stress field distribution parameters in combination with the Bayesian probability model to generate an optimized parameter set, and the optimized parameter set is fed back to the fracture dynamic modeling module for model update to make the model more accurately reflect the actual situation.

[0041] A cloud verification module, which performs multi-dimensional verification on the optimized fracture model. By constructing a reconstruction error evaluation model based on variational autoencoder, the consistency of fracture geometry and physical property parameters is verified; the Monte Carlo simulation method is combined to quantitatively analyze the model uncertainty, and finally the verification result is output and stored in the cloud database.

[0042] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1:

[0043] In the multi-source data acquisition module, a detailed description of the various types of data collected is as follows:

[0044] Seismic wave data includes the P-wave to S-wave velocity ratio, reflection coefficient matrix, and dispersion curve. The P-wave to S-wave velocity ratio reflects the differences in elastic properties of shale formations. Different formation structures and lithologies will result in different P-wave to S-wave velocity ratios. By collecting and analyzing it, the lithology changes of the formation can be understood, providing information about the characteristics of the formation medium for fracture modeling. The reflection coefficient matrix contains information on the reflection characteristics of seismic waves at different formation interfaces, which can help identify the interface positions and property differences of the formations and is of great significance for determining the possible areas where fractures may exist. The dispersion curve reflects the relationship between the frequency and propagation velocity of seismic waves, and its variation characteristics can reveal the formation structure and physical property changes, providing richer formation information for fracture modeling.

[0045] Core physical property data includes the percentage of mineral composition, Young's modulus, and fracture toughness value. The percentage of mineral composition directly affects the mechanical properties of shale and the development of fractures. Different contents of mineral components will result in different characteristics such as brittleness and toughness of shale, thus affecting the generation and propagation of fractures. Young's modulus characterizes the ability of shale to resist elastic deformation. It is an important parameter for describing the mechanical properties of shale and plays a key role in analyzing the deformation and propagation laws of fractures under stress. The fracture toughness value reflects the ability of shale to resist fracture propagation and is an important indicator for evaluating the stability of shale fractures.

[0046] Historical fracturing construction data includes the pump injection pressure curve, proppant concentration distribution, and fracturing fluid viscosity time series data. The pump injection pressure curve records the change of pressure with time during the fracturing construction. By analyzing this curve, the effect of the fracturing construction and the formation's response to pressure can be understood, and then the fracture propagation can be inferred. The proppant concentration distribution determines the effect of keeping the propped fractures open. A reasonable proppant concentration distribution can ensure that the fractures maintain a certain opening degree after fracturing, which is beneficial to oil and gas production. The fracturing fluid viscosity time series data reflects the viscosity change of the fracturing fluid during the construction process. The fracturing fluid viscosity has an important impact on the fracture propagation morphology and efficiency. By monitoring its change over time, the fracturing construction process can be better grasped. Example 2:

[0047] In the heterogeneous data fusion module, the spatio-temporal alignment algorithm is implemented according to the following steps. First, the dynamic time warping algorithm is used to perform time series matching on the drilling trajectory data and the microseismic monitoring data. The drilling trajectory data records the position and direction information of the wellbore during the drilling process, while the microseismic monitoring data reflects the time and position information of the microseismic events generated by the fracture of underground rocks. Since the acquisition time and frequency of these two types of data may be different, it is necessary to find the best time correspondence between them through the dynamic time warping algorithm. The basic idea of the dynamic time warping algorithm is to calculate the similarity between two time series and find a path to align them so that the sum of the distances of the corresponding points on the path is minimized. Suppose there are two time series and , and its distance matrix represents and The distance between them, and the cumulative distance matrix is calculated through the dynamic programming algorithm:

[0048] where , and finally the best matching path is found through backtracking to achieve the time series matching of the drilling trajectory data and the microseismic monitoring data.

[0049] Next, the Kriging interpolation method is used to complement the missing spatial data. Kriging interpolation is a spatial interpolation method based on the theory of regionalized variables, which takes into account the spatial autocorrelation of the data. Suppose there are known sampling points observations , and to estimate the value at the unknown point , the Kriging interpolation formula is: where is the weight coefficient, which is obtained by solving a system of equations. The construction of this system of equations is based on the semivariogram. The semivariogram reflects the degree of variation of spatial data. By fitting the experimental semivariogram model, such as the spherical model, the exponential model, etc., the parameters of the semivariogram are determined, and then the weight coefficient is calculated to achieve the accurate complement of the missing spatial data.

[0050] Finally, the frequency domain characteristics of the seismic wave data are extracted through Fourier transform. Fourier transform can convert the seismic wave data in the time domain into frequency domain data, and its formula is:

[0051] Among them, f(t) is the seismic wave data in the time domain, F(ω) is the data in the frequency domain, and ω is the frequency. Through Fourier transform, the frequency domain characteristics of the seismic wave data are obtained, and then a multi-source data tensor in the unified spatio-temporal coordinate system is constructed. The data after time series matching and spatial data completion are integrated together to form a unified multi-source data tensor, which facilitates subsequent data fusion and analysis. Embodiment 3:

[0052] In the crack dynamic modeling module, the crack topology generation algorithm of the graph convolutional network is carried out according to the following steps. First, the shale formation is discretized into a node graph structure, where the node attributes include lithological characteristics and stress states. The lithological characteristics include information such as the mineral composition and porosity of the shale, and the stress state includes information such as the magnitude and direction of the formation stress. The edge weights between nodes are calculated from the distance between nodes and the physical property differences. For example, for node and node , its edge weight can be calculated by the following formula:

[0053] where is the distance between node and node , and are the physical property parameters (such as porosity) of node and node respectively. Calculating the edge weights in this way can comprehensively consider the spatial distance and physical property differences between nodes and more accurately reflect the structural relationship of the formation.

[0054] Next, a multi-layer graph convolutional operator is used to extract local crack connectivity features. The graph convolutional operator can perform convolutional operations on the graph structure to extract the neighborhood features of nodes. Suppose the adjacency matrix of the graph is , the node feature matrix is , and the output of the -th layer graph convolutional operator can be calculated by the following formula:

[0055] where , is the identity matrix, is the degree matrix of , is the weight matrix of the -th layer, and is the activation function, such as the ReLU function. By stacking multi-layer graph convolutional operators, more advanced local crack connectivity features can be gradually extracted.

[0056] Then, the multi-scale topological information is fused through the skip connection mechanism. The skip connection mechanism allows the features of different layers to be directly connected, avoiding the prediction of crack propagation paths in the deep layers of the network. The gated recurrent unit (GRU) can effectively process time series data, and its update gate and reset gate are calculated as follows:

[0057] ;

[0058] where, and are weight matrices, is the hidden state at the previous moment, is the input at the current moment, is the sigmoid function. The candidate hidden state is calculated as:

[0059]

[0060] The final hidden state is calculated as:

[0061]

[0062] The generator continuously predicts the crack propagation path through the GRU according to the node states of the crack network diagram.

[0063] The discriminator uses the multi-head self-attention mechanism to calculate the crack distribution rationality score. The multi-head self-attention mechanism can learn features from different representation subspaces, and its calculation process is as follows. First, the input feature is respectively linearly transformed to obtain the query matrix , key matrix and value matrix :

[0064] ;

[0065] ;

[0066]

[0067] Then, the attention score matrix is calculated:

[0068] where, is the dimension of the key matrix . Finally, the results of multiple attention heads are concatenated through the multi-head attention mechanism:

[0069] Among them, is the output of the th attention head, is the output weight matrix. The discriminator calculates the rationality score of the fracture distribution through the multi-head self-attention mechanism and iteratively optimizes the generator parameters through the adversarial loss function, making the fracture distribution more reasonable and obtaining a more accurate dynamic model of shale fractures. Example 4:

[0070] In the parameter optimization module, a random perturbation term of the simulated annealing mechanism is introduced into the particle position update formula of the hybrid particle swarm optimization algorithm. The particle position update formula is:

[0071] Among them, represents the velocity of particle at the th iteration; represents the velocity of particle at the th iteration; represents the inertia weight, which is used to balance the global search and local search capabilities of the algorithm. Usually, it gradually decreases from a larger value during the iteration process. For example, the initial value is set to 0.9 and linearly decreases to 0.4 as the number of iterations increases; , represent the learning factors, generally taking values between [1, 2]. For example, both are taken as 1.5, which are used to adjust the step size of the particle moving towards the individual optimal position and the global optimal position; , represent two random numbers between [0, 1]; represents the historical optimal position of particle ; represents the position of particle at the th iteration; represents the global optimal position; represents the annealing coefficient, which is used to control the amplitude of the random perturbation in the simulated annealing mechanism. The initial value can be set to 0.5 and gradually decreases as the number of iterations increases; represents the cooling rate, which determines the speed at which the random perturbation weakens as the number of iterations increases. For example, the value is 0.01; represents the number of iterations; represents a random vector subject to a Gaussian distribution, which is used to enhance the global search ability of the algorithm. Its mean is 0, and the variance can be adjusted according to the actual situation.

[0072] The Bayesian probability model constructs the joint probability distribution of fracture parameters using the non-parametric kernel density estimation method. By smoothing the parameter correlation relationship in the historical fracturing data through the kernel function, assuming that the fracture parameter samples in the historical fracturing data are , the kernel function is , then the formula for kernel density estimation is:

[0073] where is the bandwidth parameter, which is determined by methods such as cross-validation. Through kernel density estimation, a conditional probability table is generated. Then, the Markov chain Monte Carlo sampling method is used to obtain the parameter optimization direction. The Markov chain Monte Carlo sampling method constructs a Markov chain whose stationary distribution is the target distribution (i.e., the joint probability distribution obtained through kernel density estimation), samples on the Markov chain to obtain a series of samples, and determines the parameter optimization direction based on the distribution of these samples, thereby realizing the multi-objective optimization of fracture permeability, porosity, and stress field distribution parameters. Example 5:

[0074] This example elaborates in detail the specific implementation methods of the reconstruction error evaluation model of the variational autoencoder, the Monte Carlo simulation method, and the model version management function in the cloud verification module. Through these methods, the optimized fracture model can be comprehensively and deeply verified, the accuracy and reliability of the model can be effectively evaluated, and the effective management and recording of the model version can be realized, which is convenient for subsequent model backtracking, comparison, and improvement, providing strong support for the practical application of shale fracture modeling.

[0075] In the cloud verification module, the reconstruction error evaluation model of the variational autoencoder is implemented as follows. The encoder uses a deep residual network to extract the fracture morphology latent variable. The deep residual network solves the problem of gradient disappearance during the training process of deep neural networks by introducing residual blocks and can better learn the features of data. Assume that the input fracture morphology data is , after a series of operations of convolutional layers and residual blocks, the output of the encoder is the latent variable . For example, a simple residual block contains two convolutional layers, and its calculation process is:

[0076] where is the input of the residual block, is the output, represents the convolution operation, is the weight of the convolutional layer. Through multiple such residual blocks, the encoder gradually extracts the high-level features of the fracture morphology and obtains the latent variable .

[0077] The decoder uses a transposed convolution network to reconstruct the three-dimensional fracture geometry. The transposed convolution network is the inverse process of the convolution network. It maps the low-dimensional latent variable back to the high-dimensional image space through transposed convolution operations. Assume the latent variable After a series of operations of deconvolution layers and activation functions, the reconstructed fracture geometry is obtained. . Taking a simple deconvolution layer as an example, its operation formula is:

[0078] where is the weight of the deconvolution layer.

[0079] The loss function contains the weighted sum of the KL divergence term and the mean square error term, which is used to quantify the model reconstruction accuracy. The loss function is calculated as:

[0080] where is the weight coefficient, whose value ranges from [0, 1], for example, taking 0.5; is the KL divergence term, which represents the difference between the approximate posterior distribution and the prior distribution , and its calculation formula is:

[0081]

[0082] In actual calculations, the reparameterization trick is usually used for approximate calculations. is the mean square error term, which is used to measure the difference between the original fracture morphology and the reconstructed fracture morphology , and the calculation formula is:

[0083] where is the number of data samples. By minimizing the loss function , the parameters of the variational autoencoder are continuously optimized, so that the reconstructed fracture geometry is as close as possible to the original structure, thereby performing consistency verification on the fracture geometry and physical property parameters.

[0084] In the Monte Carlo simulation method, a sample set of the parameter space is generated by Latin hypercube sampling. Latin hypercube sampling is a stratified sampling method, which divides the parameter space into several intervals and randomly selects sample points within each interval, so as to ensure the uniform distribution of samples in the parameter space and improve the sampling efficiency. Suppose we want to sample parameters, and each parameter has intervals, then a total of sample points are generated. For example, for these parameters of fracture permeability, porosity and stress field distribution, the value range of each parameter is divided into intervals, and a sample point is randomly selected from each interval to obtain a sample set containing 100 sample points.

[0085] Construct the mapping relationship between the crack propagation probability and the stress field perturbation based on the response surface model. The response surface model is a model that establishes the relationship between variables by fitting data. For example, a quadratic polynomial is used as the response surface model:

[0086] where, is the crack propagation probability, are the input parameters such as the stress field perturbation, , , , are the model parameters, which are estimated by methods such as the least squares method, is the error term. Through the response surface model, calculate the confidence interval and sensitivity index of the model output. The confidence interval is used to measure the uncertainty range of the model prediction results. For example, calculate the 95% confidence interval of the crack propagation probability. The sensitivity index is used to evaluate the influence degree of each input parameter on the model output. For example, use the variance decomposition method to calculate the sensitivity index :

[0087] where, represents the variance of the mean value of the model output caused by the change of parameter while fixing other parameters, is the total variance of the model output.

[0088] The cloud verification module also includes a model version management function. Generate a unique identifier for each optimized crack model through the hash algorithm. The hash algorithm maps the model data to a hash value of a fixed length. For example, use the SHA-256 hash algorithm, and its calculation formula is:

[0089] where, is the optimized crack model data, is the generated hash value. This hash value serves as the unique identifier of the model, with uniqueness and irreversibility. At the same time, record the parameter optimization path and verification results to the blockchain evidence storage system. The blockchain evidence storage system has characteristics such as decentralization and immutability. Record the parameter optimization path and verification results of the model on the blockchain to ensure the security and traceability of the data. For example, record information such as the parameter changes of each optimization, the verification time, and the verification results in the form of blockchain transactions. Each transaction contains information such as the transaction timestamp, transaction data, and the hash value of the previous transaction, forming a chain structure to facilitate the subsequent management and backtracking of the model version.

[0090] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or apparatus.

[0091] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A shale fracture modeling platform based on cloud computing, characterized in that: It includes multi-source data acquisition module, heterogeneous data fusion module, fracture dynamic modeling module, parameter optimization module and cloud verification module; The multi-source data acquisition module is used to acquire the original multi-source data set required for shale fracture modeling, specifically by deploying a distributed sensor network and geological exploration equipment to collect seismic wave data of shale formations, core physical property data, drilling trajectory data, microseismic monitoring data and historical fracturing construction data, and upload the original multi-source data set to a cloud storage cluster; The heterogeneous data fusion module is used to perform cross-domain fusion and feature enhancement on the original multi-source data, specifically to eliminate the difference in data acquisition delay through the time-space alignment algorithm, and to perform noise reduction on the seismic wave data using a multi-scale decomposition method based on wavelet transform, and to perform feature dimension reduction on the core physical property data in combination with the principal component analysis algorithm, to generate a fusion feature matrix, and to transmit the fusion feature matrix to the fracture dynamic modeling module; The fracture dynamic modeling module is used to construct a three-dimensional dynamic model of shale fractures. Specifically, a fracture topology generation algorithm based on a graph convolutional network is used to perform spatial structure modeling on the fusion feature matrix to generate an initial fracture network graph, and a generative adversarial network is introduced to dynamically optimize the fracture distribution to obtain a shale fracture dynamic model, wherein the discriminator of the generative adversarial network uses an adaptive attention mechanism to optimize weight distribution; The parameter optimization module is used to iteratively optimize the physical property parameters of the fracture model, specifically using a hybrid particle swarm optimization algorithm combined with a Bayesian probability model to perform multi-objective optimization on fracture permeability, porosity and stress field distribution parameters, generate an optimized parameter set, and feed the optimized parameter set back to the fracture dynamic modeling module for model update; The cloud verification module is used to perform multi-dimensional verification on the optimized crack model. Specifically, it constructs a reconstruction error evaluation model based on a variational autoencoder to perform consistency verification on the crack geometry and physical property parameters. At the same time, it combines the Monte Carlo simulation method to quantify the model uncertainty, output the verification results and store them in the cloud database.

2. The cloud computing-based shale fracture modeling platform according to claim 1, characterized in that: In the multi-source data acquisition module, the seismic wave data include the ratio of longitudinal and transverse wave velocities, the reflection coefficient matrix and the dispersion curve; the core physical property data include the mineral composition percentage, Young's modulus and fracture toughness value; the historical fracturing construction data include the pumping pressure curve, the proppant concentration distribution and the fracturing fluid viscosity time series data.

3. The cloud computing-based shale fracture modeling platform according to claim 2, characterized in that: In the heterogeneous data fusion module, the spatiotemporal alignment algorithm specifically includes the following steps: time series matching of drilling trajectory data and microseismic monitoring data is performed based on a dynamic time warping algorithm, the Kriging interpolation method is used to complete the missing spatial data, and the frequency domain characteristics of the seismic wave data are extracted through Fourier transform to construct a multi-source data tensor in a unified spatiotemporal coordinate system.

4. The cloud computing-based shale fracture modeling platform according to claim 3, characterized in that: In the fracture dynamic modeling module, the fracture topology generation algorithm of the graph convolutional network specifically includes: discretizing the shale formation into a node graph structure, in which the node attributes include lithological characteristics and stress state, and the edge weight is calculated by the distance between nodes and the physical property difference; using a multi-layer graph convolution operator to extract local fracture connectivity characteristics, and fusing multi-scale topological information through a jump connection mechanism to generate a weighted fracture network graph.

5. The cloud computing-based shale fracture modeling platform according to claim 4, characterized in that: The dynamic optimization process of the generative adversarial network includes: the generator adopts a sequence generation model with a gated recurrent unit to predict the crack extension path according to the node status of the crack network graph; the discriminator uses a multi-head self-attention mechanism to calculate the rationality score of the crack distribution, and iteratively optimizes the generator parameters through an adversarial loss function.

6. The cloud computing-based shale fracture modeling platform according to claim 5, characterized in that: In the hybrid particle swarm optimization algorithm, the particle position update formula introduces the random perturbation term of the simulated annealing mechanism, specifically: in, Indicates At the iteration, the particle speed; Indicates At the iteration, the particle speed; Represents the inertia weight, which is used to balance the global search and local search capabilities of the algorithm; , represents the learning factor, which is used to adjust the step size of the particle moving to the individual optimal position and the global optimal position; , Represents two random numbers between [0, 1]; Represents particles The best historical position; Indicates At the iteration, the particle location; represents the global optimal position; represents the annealing coefficient, which is used to control the amplitude of random perturbations in the simulated annealing mechanism; represents the cooling rate, which determines how fast the random perturbation weakens as the number of iterations increases; Indicates the number of iterations; Represents a random vector that follows a Gaussian distribution, which is used to enhance the global search capability of the algorithm.

7. The cloud computing-based shale fracture modeling platform according to claim 1, characterized in that: The Bayesian probability model uses a non-parametric kernel density estimation method to construct a joint probability distribution of fracture parameters, specifically, by smoothing the parameter association relationship in historical fracturing data through a kernel function, generating a conditional probability table, and using a Markov chain Monte Carlo sampling method to obtain a parameter optimization direction.

8. The cloud computing-based shale fracture modeling platform according to claim 1, characterized in that: In the reconstruction error evaluation model of the variational autoencoder, the encoder uses a deep residual network to extract latent variables of crack morphology, the decoder uses a deconvolution network to reconstruct the three-dimensional crack geometric structure, and the loss function includes a weighted sum of a KL divergence term and a mean square error term, which is used to quantify the model reconstruction accuracy.

9. The cloud computing-based shale fracture modeling platform according to claim 1, characterized in that: In the Monte Carlo simulation method, a parameter space sample set is generated by Latin hypercube sampling, and a mapping relationship between crack extension probability and stress field disturbance is constructed based on a response surface model to calculate the confidence interval and sensitivity index of the model output.

10. The cloud computing-based shale fracture modeling platform according to claim 1, characterized in that: The cloud verification module also includes a model version management function, which specifically generates a unique identifier for each optimized crack model through a hash algorithm, and records the parameter optimization path and verification results to the blockchain evidence storage system.

Images