A permeability prediction method and system based on multi-source attribute coupling

Abstract

The present invention discloses a permeability prediction method and system based on multi-source attribute coupling, comprising: obtaining multi-source attribute information and permeability data, wherein the multi-source attribute information includes fracture density, porosity and water inflow; constructing a multi-source attribute-permeability coupling prediction model, and training the multi-source attribute-permeability coupling prediction model based on the multi-source attribute information and permeability data; inputting the multi-source attribute information into the trained multi-source attribute-permeability coupling prediction model to obtain predicted permeability; the present invention takes the coupling relationship between multi-attribute information such as fracture density, porosity, and water inflow and permeability as the core of the standardized model, adopts mathematical formulas to describe the coupling relationship between each attribute, and combines experimental and numerical simulation data to establish its physical constraints in the deep learning model.

Description

Technical Field

[0001] The present invention relates to the fields of underground engineering, geotechnical engineering and geological exploration, and in particular to a permeability prediction method and system based on multi-source attribute coupling. Background Art

[0002] Permeability is a key parameter in geotechnical engineering and groundwater resource development. It reflects the ability of a geotechnical medium to accommodate fluid flow. Permeability directly impacts groundwater flow, grouting effectiveness, contaminant diffusion, and oil and gas recovery from underground reservoirs. Therefore, accurately understanding the permeability of geological structures is crucial for engineering design and implementation. In geotechnical engineering, permeability is widely used in hydraulic structure construction, underground engineering design, and foundation pit support. For example, during the construction of infrastructure such as tunnels and basements, permeability helps assess groundwater level fluctuations and flow conditions, ensuring the waterproofness and safety of the structure. In grouting, permeability determines the effectiveness of grout diffusion, impacting the reinforcement and stability of underground structures. Understanding formation permeability allows engineers to optimize grouting designs, ensuring that the grout effectively penetrates and fills pores, enhancing the bearing capacity and durability of the structure. Therefore, permeability is not only a crucial physical parameter in geotechnical engineering but also a guiding factor in a variety of underground and environmental engineering projects.

[0003] In underground engineering and geological research, property parameters (fracture density, porosity, and water inflow) and permeability are important indicators for describing the structure and physical properties of geological bodies. Drilling and laboratory testing can obtain property and permeability data for selected areas. However, due to the complexity of the strata and limited data, these properties often vary spatially, and permeability information lacks comprehensive coverage, leading to significant errors in modeling underground structures. Furthermore, existing permeability prediction models suffer from low prediction accuracy and large errors, necessitating further improvement and optimization. By accurately acquiring and optimizing permeability models, engineers can better assess the flow characteristics of underground media, providing a scientific basis for project design and implementation, and thus improving the safety, cost-effectiveness, and sustainability of projects. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a permeability prediction method and system based on multi-source attribute coupling. Based on the joint iterative reconstruction of multi-source information, comprehensive and dynamic coupling of multiple attributes and permeability in space is achieved to improve the prediction accuracy and adaptability of the model.

[0005] The technical solutions of the present invention are as follows:

[0006] In a first aspect of the present invention, a permeability prediction method based on multi-source attribute coupling is provided, comprising:

[0007] Acquiring multi-source attribute information to be predicted and preprocessing the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow;

[0008] The processed multi-source attribute information is input into the trained multi-source attribute-permeability coupling prediction model to obtain the predicted permeability;

[0009] Among them, the training process of the multi-source attribute-permeability coupling prediction model is as follows: historical multi-source attribute information and permeability data are obtained, and the multi-source attribute information is processed using a numerical interpolation method; then, based on the historical multi-source attribute information and permeability data, a deep learning prediction model that combines physical constraints and multi-task learning is trained to obtain a complex relationship model of multi-attribute-permeability; finally, the complex relationship model of multi-attribute-permeability is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model.

[0010] In some embodiments of the present invention, the multi-round joint iterative reconstruction process specifically includes: first, using known experimental and numerical simulation data to pre-train the complex relationship model of multiple attributes and permeability, so that the model can initially learn the coupling relationship between multiple attributes and permeability; then, each time the newly collected drilling exposure information is added to the training data, reconstruction is performed in an iterative manner, and the model uses the new data as the input for the next round of training, and continuously updates the model parameters according to the new physical data constraints.

[0011] In some embodiments of the present invention, the multi-source attribute-permeability coupling prediction model includes an input layer, a feature extraction layer, a coupling relationship layer and a fusion layer, wherein the coupling relationship layer introduces a multi-task learning module, passes the input multi-attribute information into three parallel branch models, and outputs fracture permeability, pore permeability and water permeability respectively.

[0012] In some embodiments of the present invention, the three parallel branch models include: a coupling model of fracture density and permeability, a coupling model of porosity and permeability, and a coupling model of water inflow and permeability;

[0013] The coupling model of fracture density and permeability is:

[0014]

[0015] Where k f represents the permeability under fracture control; D f is the crack density; a and b are constants determined by experiments;

[0016] The coupling model of porosity and permeability is:

[0017]

[0018] Where k p represents the permeability under pore control; φ is the porosity; S is the specific surface area; C is a constant;

[0019] The coupling model of water inflow and permeability is:

[0020]

[0021] Where Q is the water inflow; k w is the permeability under water inflow control; A is the flow cross-sectional area; Δh is the head difference; L is the flow path length.

[0022] In some embodiments of the present invention, the fusion layer performs weighted summation of the fracture permeability, pore permeability and water inflow permeability obtained by the coupling relationship layer, and outputs a comprehensive permeability as the predicted permeability.

[0023] In some embodiments of the present invention, a physical constraint layer is further included. The physical constraint layer adds a physical constraint term to the loss function. The loss function is defined as follows:

[0024]

[0025] in: is the mean square error between the model prediction and the true permeability; is the regularization loss term;

[0026] Physical consistency loss based on multi-attribute coupling formula

[0027]

[0028] where k f 、k p and k w It is calculated from the branch output of the model; k is the comprehensive penetration rate; λ1 and λ2 are weight hyperparameters used to balance the impact of different loss terms; α, β, and γ are the weights of the factors.

[0029] In some embodiments of the present invention, the training and optimization process of the multi-attribute-permeability complex relationship model is as follows:

[0030] Initialize model parameters and weights: preset model hyperparameters, such as weights λ1 and λ2 in the loss function, and initialize network parameters;

[0031] Batch training: Update network parameters based on the backpropagation algorithm, and calculate the loss value and gradient through forward propagation and backpropagation of each small batch of data;

[0032] Physical constraint adjustment: In each iteration, the values ​​of α, β, and γ are adjusted so that the comprehensive permeability k output by the model satisfies the physical constraint relationship as much as possible, and the physical consistency is gradually optimized;

[0033] Evaluation and tuning: Evaluate the model's generalization ability and physical consistency through cross-validation and loss curve analysis, and adjust the model's hyperparameters such as the number of layers and nodes if necessary.

[0034] In some embodiments of the present invention, the numerical interpolation method uses a radial basis function interpolation method to expand the spatial coverage of multi-source attribute information.

[0035] In some embodiments of the present invention, the sources of the permeability data include drilling test data, laboratory test data, and numerical simulation data; the training of the deep learning model using the multi-source attribute information and permeability data includes coupling data from different sources, and using adjacent attributes as a guide to obtain a complex multi-attribute-permeability relationship model through repeated training.

[0036] In a second aspect of the present invention, a permeability prediction system based on multi-source attribute coupling is provided, comprising:

[0037] A data processing module is configured to: obtain multi-source attribute information to be predicted and pre-process the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow;

[0038] The prediction module is configured to: input the processed multi-source attribute information into the trained multi-source attribute-permeability coupling prediction model to perform prediction and obtain predicted permeability;

[0039] Among them, the training process of the multi-source attribute-permeability coupling prediction model is as follows: first, the multi-source attribute information is processed by a numerical interpolation method; then, based on the multi-source attribute information and permeability data, a deep learning prediction model combining physical constraints and multi-task learning is trained to obtain a complex relationship model of multi-attribute-permeability; finally, the complex relationship model of multi-attribute-permeability is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model.

[0040] One or more technical solutions of the present invention have the following beneficial effects:

[0041] (1) The permeability prediction model adopted by the permeability prediction method provided by the present invention takes the coupling relationship between multi-attribute information such as crack density, porosity, and water inflow and permeability as the core of the standardized model in the coupling process of underground medium properties, adopts mathematical formulas to describe the coupling relationship between various attributes, and combines experimental and numerical simulation data to establish its physical constraints in the deep learning model, thereby improving the accuracy of the permeability prediction results. In addition, a physical constraint layer and a physical consistency loss function based on the multi-attribute coupling formula are adopted to ensure that the model not only considers the prediction error during the optimization process, but also meets the requirements of physical laws. Through physical constraints, the output of the model is more consistent with the permeability characteristics under actual geological conditions, avoiding the model from deviating from the actual physical process, thereby enhancing the credibility and reliability of the prediction results.

[0042] (2) This paper establishes a more accurate and dynamically adaptable permeability prediction model by deeply coupling multi-source attribute information (such as fracture density, porosity, and water inflow) with permeability data and combining it with a training strategy of multiple rounds of joint iterative reconstruction. Compared with traditional models based on a single data source, this method can comprehensively consider the complex relationships between different attributes, thereby significantly improving the accuracy and stability of permeability prediction and adapting to different geological conditions and engineering requirements.

[0043] (3) The permeability prediction model used in the permeability prediction method provided by the present invention is automatically updated and optimized each time new data (such as drill hole information) is acquired through multiple rounds of joint iterative reconstruction, continuously improving the accuracy of predictions. This dynamic iterative mechanism enables the model to reflect changes in geological structure in real time over long periods of time. In particular, during dynamic monitoring and engineering implementation, the prediction results can be adjusted based on real-time data, enhancing the model's flexibility and real-time performance, and adapting to the ever-changing and complex conditions in the engineering environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Flowchart of the permeability prediction method based on multi-source attribute coupling of the present invention;

[0045] Figure 2 This is a structural diagram of the multi-source attribute-permeability coupling prediction model of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Example 1

[0048] In a typical embodiment of the present invention, a permeability prediction method based on multi-source attribute coupling is proposed, such as Figure 1 Shown, including:

[0049] Input the multi-source attribute information to be predicted into the trained multi-source attribute-permeability coupling prediction model to obtain the predicted permeability;

[0050] Step 1: obtaining multi-source attribute information to be predicted and preprocessing the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow;

[0051] Among them, a numerical interpolation method is used to process multi-source attribute information; the numerical interpolation method uses a radial basis function interpolation method to expand the spatial coverage of the multi-source attribute information, and uses a numerical model to perform attribute modeling and interpolation analysis, which can fill the missing attribute data caused by uneven borehole distribution. The interpolation method can use a radial basis function interpolation method or other suitable numerical interpolation methods to ensure the smoothness and consistency of the attribute data.

[0052] Specifically, the processing process includes:

[0053] First, the input multi-source attribute data (such as fracture density, porosity, and water inflow) is gridded based on the geometric scope and resolution requirements of the study area. Using background grid generation techniques, the study area is decomposed into uniform or adaptive three-dimensional grid cells. The attribute values ​​of each grid cell are derived from existing data points using data interpolation techniques (such as cubic spline interpolation and radial basis function interpolation), ensuring the spatial continuity and integrity of the attribute data.

[0054] Attribute interpolation and mapping:

[0055] For attribute data in unsampled areas, interpolation algorithms are used to calculate attribute values ​​at unknown points within the grid. For example, for fracture density, radial basis function interpolation is used to calculate the density distribution of surrounding grid points from known fracture density points. For porosity, a weighted average method based on volume fraction is used for interpolation. For water inflow, time series analysis is used to predict water inflow properties at unsampled points. This approach effectively addresses data sparsity while maintaining spatial consistency of physical properties.

[0056] Random point sampling:

[0057] After completing the grid division and attribute interpolation, a certain number of representative sampling points are randomly selected from the grid cells as the input data of the deep learning model. These points can be selected according to the following principles:

[0058] Uniformity: Through spatially uniform sampling, ensure that the selected points can fully cover the distribution of different attributes in the study area.

[0059] Representativeness: Based on the gradient changes of attribute data (such as fracture density gradient or porosity gradient), the density of sampling points is increased in areas with large attribute changes to improve the modeling accuracy of key areas.

[0060] Randomness: Based on uniformity and representativeness, random perturbations are added to ensure the diversity of input data and prevent model overfitting.

[0061] Feature preprocessing:

[0062] Normalize or standardize the attribute values ​​of randomly selected sampling points to unify the dimensions and value ranges of different attributes. Furthermore, design specialized encoding methods for different data types (such as scalar and image attributes) to adapt them to the input format of deep learning models. For example, convert the water inflow attribute into a time series input, and convert the fracture density distribution into a two-dimensional image feature.

[0063] Batch generation:

[0064] Finally, the data of the sampling points are organized into batches, each batch contains a certain number of sample points, and the order of the data within the batch is randomly adjusted to enhance the robustness and generalization ability of model training.

[0065] Step 2: Input the processed multi-source attribute information into the trained multi-source attribute-permeability coupling prediction model to obtain the predicted permeability;

[0066] Among them, the training process of the multi-source attribute-permeability coupling prediction model is as follows: historical multi-source attribute information and permeability data are obtained, and the multi-source attribute information is processed using a numerical interpolation method; then, based on the historical multi-source attribute information and permeability data, a deep learning prediction model that combines physical constraints and multi-task learning is trained to obtain a complex relationship model of multi-attribute-permeability; finally, the complex relationship model of multi-attribute-permeability is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model.

[0067] Specifically, a deep neural network model combining multi-task learning and physical constraints is constructed as a multi-source attribute-permeability coupling prediction model, such as Figure 2 As shown, it includes input module, feature extraction module, coupling relationship module, fusion module and physical constraint module.

[0068] Input Module: Receives multi-source attribute information, including parameters such as fracture density, porosity, and water inflow, and inputs them by attribute type. Each input data point can contain multiple spatiotemporal observation points and features dynamic updates, supporting real-time data streaming.

[0069] Feature extraction module: Uses multiple convolutional layers or multi-layer perceptrons (MLPs) to extract complex nonlinear relationships between attributes.

[0070] Coupling relationship module: introduces a multi-task learning module, which inputs multi-attribute data into three parallel branches and outputs the fracture permeability k f , pore permeability k p , water permeability k w ; These branches share some network layers, but have independent weights in feature extraction and fusion to capture the permeability characteristics of different attributes.

[0071] The three parallel branch models include: a coupling model of fracture density and permeability, a coupling model of porosity and permeability, and a coupling model of water inflow and permeability;

[0072] The coupling model of fracture density and permeability is:

[0073]

[0074] Where k f represents the permeability under fracture control; D f is the fracture density; a and b are constants determined experimentally. This formula shows that permeability increases nonlinearly with increasing fracture density. Based on experimental and numerical simulation data, the values ​​of constants a and b can be determined to make the model more realistic.

[0075] The coupling model of porosity and permeability is:

[0076]

[0077] Where k p represents the permeability under pore control; φ is the porosity; S is the specific surface area; C is a constant; this model reflects the nonlinear relationship between porosity and permeability, and specific values ​​can be obtained through experiments to adapt to different pore characteristics.

[0078] The coupling model of water inflow and permeability is:

[0079]

[0080] Where Q is the water inflow; k w is the permeability under water inflow control; A is the flow cross-sectional area; Δh is the head difference; L is the flow path length; the relationship between water inflow and permeability can be described according to Darcy's law, and the permeability k can be calculated by measuring the water inflow in the experiment. w , thus constructing the coupling standard of permeability and water inflow.

[0081] Fusion module: Based on the multi-attribute-permeability coupling formula, the three permeabilities are weighted and summed to output the comprehensive permeability k;

[0082] Time Series Modeling Module: Based on a time series convolutional network, it captures the temporal variation patterns and characterizes the dynamic behavior of the slurry diffusion process. Combined with the attention mechanism, it assigns higher weights to key moments or regions, optimizing the extraction of spatiotemporal features.

[0083] The main functions of the spatiotemporal modeling module include:

[0084] (1) Time feature capture: Track the slurry diffusion trend at different time points and predict the permeability changes at future moments.

[0085] (2) Spatial feature modeling: combining geological data with drilling layout to simulate the distribution characteristics of slurry diffusion in space; unifying the temporal and spatial characteristics through the time-space fusion layer to capture the “time-space” coupling characteristics of the slurry diffusion process.

[0086] (3) Dynamic weight adjustment: Combined with the attention mechanism, key areas and key moments are given higher weights in the spatiotemporal dimension. For example, the weights of grid cells near the grouting point and the initial diffusion time point are strengthened to ensure that the model accurately depicts the initial diffusion behavior and the spatial permeability gradient change.

[0087] (4) Physical constraint embedding: Considering the physical constraints of the slurry, the constraints are directly embedded into the space-time module to ensure that the prediction results conform to the actual engineering physics laws.

[0088] Physical consistency embedding module: Add physical constraints to the loss function to ensure that the model meets the constraints of physical formulas as much as possible during training.

[0089] To optimize the multi-source attribute-permeability coupled prediction model based on joint iterative reconstruction, deep learning models can use physical constraints as a supervision mechanism to embed the relationship between multi-source attributes (fractures, pores, water inflow) and permeability into the model to ensure physical consistency during training.

[0090] Adding a regularization term based on physical constraints to the loss function allows the model to not only minimize the prediction error but also follow the physical constraint formula. The loss function is defined as follows:

[0091]

[0092] in: is the mean square error (MSE) between the model prediction and the true permeability; Physical consistency loss based on multi-attribute coupling formulation; It is a regularization loss term used to guide the model to follow certain additional structures or constraints during training to ensure the stability and simplicity of the model or to follow certain prior assumptions.

[0093]

[0094] Among them, k f 、k p and k w It is calculated from the branch output of the model; k is the comprehensive penetration rate; λ1 and λ2 are weight hyperparameters used to balance the impact of different loss terms; α, β, and γ are the weights of the factors.

[0095] The weight factors (α, β, γ and λ1, λ2) are dynamically adjusted during the training process to ensure a balance between physical consistency and prediction accuracy.

[0096] The above model structure has made the following improvements on the existing neural network structure:

[0097] Embedded physical constraints: To ensure physical consistency during computation, the model embeds physical models and boundary conditions related to slurry diffusion during training. Specifically, the neural network model's loss function not only considers prediction errors but also incorporates errors from the physical model, making the output more consistent with actual physical laws.

[0098] Multi-source data fusion: By introducing multi-source data (such as resistivity, radar signals, and drill hole data), the model can achieve joint learning of different data types. This process requires a special fusion module to achieve synergy between different data.

[0099] Spatiotemporal feature learning: Focusing on the diffusion of slurry in complex geological environments, the model is specially designed with a structure that can handle spatiotemporal features to ensure accurate capture of slurry diffusion at different time points.

[0100] Dynamic training and updating mechanism: To cope with the dynamic changes in field monitoring data, the model is designed with a mechanism that can be continuously updated and adjusted to ensure effective predictions at different work stages (such as before grouting, after grouting, and re-drilling).

[0101] Specifically, the model iteratively updates and trains data according to the following process to improve its adaptability and accuracy to different geological conditions.

[0102] 1. Initial data training: First, the model is pre-trained using known experimental and numerical simulation data to enable the model to initially learn the coupling relationship between multiple attributes and permeability.

[0103] Specifically, they include:

[0104] (1) A numerical interpolation method is used to process multi-source attribute information; the numerical interpolation method uses a radial basis function interpolation method to expand the spatial coverage of the multi-source attribute information, uses a numerical model to perform attribute modeling and interpolation analysis, and can fill the missing attribute data caused by uneven borehole distribution. The interpolation method can use a radial basis function interpolation method or other appropriate numerical interpolation method to ensure the smoothness and consistency of the attribute data.

[0105] (2) A deep learning model is trained based on multi-source attribute and permeability data to obtain a complex multi-attribute-permeability relationship model; the multi-source attribute information and permeability data training of the deep learning model includes coupling data from different sources and using adjacent attributes as a guide to obtain a complex multi-attribute-permeability relationship model through repeated training. The deep learning model can use a convolutional neural network or other network structure suitable for geological attribute representation to ensure that it can capture the spatial nonlinear relationship between geological attributes.

[0106] Coupling data from different sources involves integrating and correlating multi-source attribute data (such as fracture density, porosity, and water inflow) with permeability data to facilitate deep learning model training. The core of coupling is to fuse these different types of data into a unified input format, enabling the deep learning model to simultaneously consider the interactions between these different attributes and their relationship to permeability during the learning process.

[0107] Coupling data from different sources includes:

[0108] Data splicing: Attribute data from different sources (such as drill hole tests, laboratory measurements, numerical simulations, etc.) are spliced ​​into a multidimensional vector and used as input to the model. This attribute information is combined into a unified data structure through splicing for deep learning model processing.

[0109] Weighted fusion: assign different weights to attribute data from different sources, and perform weighted averaging based on the reliability or importance of different data to form a comprehensive input data.

[0110] Feature interaction: Through neural network layers or convolution operations in deep learning, the relationships and interaction features between different attributes are automatically learned. For example, convolutional neural networks (CNNs) can capture the spatial relationships between geological attributes through convolution operations.

[0111] The "adjacent attributes as guidance" in this process means that during the training process, the model will pay attention to adjacent attributes (such as the relationship between water content and permeability) and optimize the model through repeated training to better capture the spatial nonlinear relationship between attributes.

[0112] 2. Joint iterative reconstruction: Each newly acquired drill hole exposure information is added to the training data and reconstructed in an iterative manner. The model uses the new data as the input for the next round of training and continuously updates the model parameters based on the new physical data constraints to optimize the prediction accuracy of the coupled model.

[0113] Specifically, after constructing a deep learning model coupling relationship using experimental and numerical simulation data, the model is iteratively updated. In each iteration, the latest drill hole information (i.e., field data) is used as new data input. The deep learning model continuously optimizes the multi-attribute-permeability coupling relationship, gradually refining the model's accuracy and applicability.

[0114] Based on multiple rounds of joint iterative reconstruction, a multi-source attribute-permeability coupling prediction model that meets engineering needs is formed and refined into a standardized form of the model, so that it can be quickly adapted to different underground engineering applications and realize the rapid integration and modeling of multiple attribute and permeability information.

[0115] 3. Multi-source data fusion: The model gradually accumulates more collected data (such as laboratory data, numerical simulation data, and drill hole exposure data), and performs multi-source data normalization to ensure the continuity and consistency between new data and old data, and avoid model deviations caused by differences in data from different sources.

[0116] Adding physical constraints to the above model for optimization training includes:

[0117] 1. Initialize model parameters and weights: Preset the model's hyperparameters, such as the weights λ1 and λ2 in the loss function, and initialize the network parameters.

[0118] 2. Batch training: Update network parameters based on the backpropagation algorithm, and calculate the loss value and gradient through forward propagation and backpropagation of each small batch of data.

[0119] 3. Physical constraint adjustment: In each iteration, the values ​​of α, β, and γ are adjusted so that the comprehensive permeability k output by the model satisfies the physical constraint relationship as much as possible, and the physical consistency is gradually optimized.

[0120] 4. Evaluation and Tuning: Evaluate the model's generalization ability and physical consistency through cross-validation and loss curve analysis. If necessary, adjust hyperparameters such as the number of model layers and nodes.

[0121] After training, the model automatically updates its comprehensive permeability predictions based on newly acquired multi-source data, enabling real-time permeability monitoring and dynamic control of underground projects. The comprehensive permeability k output by the model can be directly used for engineering predictions and provide guidance for drilling layout, seepage control, and other aspects.

[0122] Example 2

[0123] In a typical embodiment of the present invention, a permeability prediction method based on multi-source attribute coupling is provided, comprising:

[0124] A data processing module is configured to: obtain multi-source attribute information to be predicted and pre-process the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow;

[0125] The prediction module is configured to: input the processed multi-source attribute information into the trained multi-source attribute-permeability coupling prediction model to perform prediction and obtain predicted permeability;

[0126] Among them, the training process of the multi-source attribute-permeability coupling prediction model is as follows: historical multi-source attribute information and permeability data are obtained, and the multi-source attribute information is processed using a numerical interpolation method; then, based on the historical multi-source attribute information and permeability data, a deep learning prediction model that combines physical constraints and multi-task learning is trained to obtain a complex relationship model of multi-attribute-permeability; finally, the complex relationship model of multi-attribute-permeability is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model.

[0127] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A permeability prediction method based on multi-source attribute coupling, characterized in that: include: Acquiring multi-source attribute information to be predicted and preprocessing the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow; The processed multi-source attribute information is input into the trained multi-source attribute-permeability coupling prediction model to obtain the predicted permeability; The training process of the multi-source attribute-permeability coupling prediction model is as follows: historical multi-source attribute information and permeability data are obtained, and the multi-source attribute information is processed using a numerical interpolation method; then, a deep learning prediction model that combines physical constraints and multi-task learning is trained based on the historical multi-source attribute information and permeability data to obtain a complex multi-attribute-permeability relationship model; finally, the complex multi-attribute-permeability relationship model is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model; The multi-source attribute-permeability coupling prediction model includes an input layer, a feature extraction layer, a coupling relationship layer, and a fusion layer. The coupling relationship layer introduces a multi-task learning module to pass the input multi-attribute information into three parallel branch models, and outputs fracture permeability, pore permeability, and water permeability respectively. The three parallel branch models include: a coupling model of fracture density and permeability, a coupling model of porosity and permeability, and a coupling model of water inflow and permeability; The coupling model of fracture density and permeability is: Where, represents the permeability under fracture control; is the crack density; and is a constant determined experimentally; The coupling model of porosity and permeability is: Where, represents the permeability under pore control; is the porosity; is the specific surface area; is a constant; The coupling model of water inflow and permeability is: Where, is the water inflow; is the permeability under water inflow control; is the flow cross-sectional area; is the water head difference; is the flow path length.

2. The permeability prediction method based on multi-source attribute coupling according to claim 1, characterized in that: The multi-round joint iterative reconstruction process specifically includes: first, using known experimental and numerical simulation data to pre-train the complex relationship model between multiple attributes and permeability, so that the model can initially learn the coupling relationship between multiple attributes and permeability; then, each time the newly collected drilling exposure information is added to the training data, reconstruction is performed in an iterative manner. The model uses the new data as the input for the next round of training and continuously updates the model parameters according to the new physical data constraints.

3. The permeability prediction method based on multi-source attribute coupling according to claim 1, characterized in that: The fusion layer performs weighted summation of the fracture permeability, pore permeability and water inflow permeability obtained from the coupling relationship layer, and outputs a comprehensive permeability as the predicted permeability.

4. The permeability prediction method based on multi-source attribute coupling according to claim 1, characterized in that: The multi-source attribute-permeability coupling prediction model also includes a physical constraint layer, which adds a physical constraint term to the loss function. The loss function is defined as follows: in: is the mean square error between the model prediction and the true permeability; is the regularization loss term; Physical consistency loss based on multi-attribute coupling formula: in 、 and Calculated by the branch output of the model; is the comprehensive penetration rate; and is a weight hyperparameter used to balance the effects of different loss terms; , β, and γ are the weights of the factors.

5. The permeability prediction method based on multi-source attribute coupling according to claim 4, characterized in that: The training and optimization process of the multi-attribute-permeability complex relationship model is as follows: Initialize model parameters and weights: preset model hyperparameters, such as weights in the loss function 、 , and initialize network parameters; Batch training: Update network parameters based on the backpropagation algorithm, and calculate the loss value and gradient through forward propagation and backpropagation of each small batch of data; Physical Constraint Adjustment: In each iteration, adjust , β, γ values, so that the comprehensive permeability output by the model Satisfy physical constraints as much as possible and gradually optimize physical consistency; Evaluation and tuning: Evaluate the model's generalization ability and physical consistency through cross-validation and loss curve analysis, and adjust the model's hyperparameters such as the number of layers and nodes if necessary.

6. The permeability prediction method based on multi-source attribute coupling according to claim 1, characterized in that: The numerical interpolation method uses a radial basis function interpolation method to expand the spatial coverage of multi-source attribute information.

7. The permeability prediction method based on multi-source attribute coupling according to claim 1, characterized in that: The sources of the permeability data include drilling test data, laboratory test data and numerical simulation data; the training of the deep learning model using the multi-source attribute information and permeability data includes coupling data from different sources and using adjacent attributes as a guide to obtain a complex multi-attribute-permeability relationship model through repeated training.

8. A permeability prediction method based on multi-source attribute coupling, characterized in that: include: A data processing module is configured to: obtain multi-source attribute information to be predicted and pre-process the multi-source attribute information, wherein the multi-source attribute information includes fracture density, porosity, and water inflow; The prediction module is configured to: input the processed multi-source attribute information into the trained multi-source attribute-permeability coupling prediction model to perform prediction and obtain predicted permeability; The training process of the multi-source attribute-permeability coupling prediction model is as follows: historical multi-source attribute information and permeability data are obtained, and the multi-source attribute information is processed using a numerical interpolation method; then, a deep learning prediction model that combines physical constraints and multi-task learning is trained based on the historical multi-source attribute information and permeability data to obtain a complex multi-attribute-permeability relationship model; finally, the complex multi-attribute-permeability relationship model is reconstructed through multiple rounds of joint iterative reconstruction to obtain a multi-source attribute-permeability coupling prediction model; The multi-source attribute-permeability coupling prediction model includes an input layer, a feature extraction layer, a coupling relationship layer, and a fusion layer. The coupling relationship layer introduces a multi-task learning module to pass the input multi-attribute information into three parallel branch models, and outputs fracture permeability, pore permeability, and water permeability respectively. The three parallel branch models include: a coupling model of fracture density and permeability, a coupling model of porosity and permeability, and a coupling model of water inflow and permeability; The coupling model of fracture density and permeability is: Where, represents the permeability under fracture control; is the crack density; and is a constant determined experimentally; The coupling model of porosity and permeability is: Where, represents the permeability under pore control; is the porosity; is the specific surface area; is a constant; The coupling model of water inflow and permeability is: Where, is the water inflow; is the permeability under water inflow control; is the flow cross-sectional area; is the water head difference; is the flow path length.

Images