A Deep Learning-Based Prediction Method for Multi-Level Branch Joint Prop Placement and Diversion Capacity

Abstract

This invention relates to a deep learning-based prediction method for the joint proppant placement and conductivity of multi-level branch fractures, belonging to the field of petroleum technology. The method includes sampling and combining branch fracture parameters, construction parameters, and reservoir parameters to generate multiple simulation schemes covering various working conditions. Each simulation scheme includes a combination of branch fracture parameters, construction parameters, and reservoir parameters. The generated simulation schemes are input into a pre-constructed numerical model covering multi-level branch fracture proppant migration and placement to simulate and obtain two-dimensional matrices of proppant placement and conductivity at each time step, forming a time-series training dataset. A parameterized multi-task U-shaped convolutional neural network is constructed, comprising a one-dimensional vector input layer, a two-dimensional matrix input layer, a CNN shared feature extraction module, a time-series module, and a dual-U Net output module.

Description

Technical Field

[0001] This invention relates to the field of petroleum technology, and in particular to a method for predicting the combined proppant placement and flow capacity of multi-level branched fractures based on deep learning. Background Technology

[0002] With the continuous advancement of unconventional oil and gas resource development, low-permeability tight oil and gas reservoirs, shale oil and gas, and coalbed methane have become important energy development targets. Due to the characteristics of low porosity, low permeability, and low natural productivity of these reservoirs, hydraulic fracturing technology is usually required to modify the reservoirs to form complex fracture networks and improve fracture conductivity, thereby enhancing reservoir productivity.

[0003] During fracturing operations, the proppant placement morphology and fracture conductivity within the fracture are crucial factors determining the fracturing effect. The migration, deposition, and distribution of proppant in the main and branch fractures not only affect the effective propping range but also directly influence the formation and retention of flow channels. Therefore, accurately predicting the proppant placement morphology and conductivity distribution in complex fracture networks is of great significance for optimizing fracturing schemes, adjusting construction parameters, and evaluating results.

[0004] However, under multi-level branched fracture network conditions, the transport, diversion, and deposition processes of proppant between the main fracture and branch fractures are influenced by a variety of factors, including fracture geometry, construction parameters, and reservoir conditions, exhibiting significant complexity and stage-specific variations. Meanwhile, the formation of fracture conductivity is closely related to the proppant placement state, and the two influence each other; therefore, rapid and accurate prediction of these processes has always been a technical challenge in this field.

[0005] Existing methods mainly include physical experiments, analytical models, and numerical simulations. While physical experiments can intuitively reflect the proppant placement process, they suffer from high experimental costs, difficulties in model construction, and limitations in scale. Analytical models are computationally efficient, but they are usually based on simplifying assumptions and are difficult to describe the complex changes in complex fracture networks. Numerical simulation methods can characterize the proppant migration process with greater precision, but they are computationally intensive and time-consuming, making it difficult to meet the needs of rapid on-site prediction and real-time optimization.

[0006] In recent years, deep learning methods have begun to be applied to proppant placement or conductivity prediction. However, existing methods are mostly single-task static predictions, typically estimating only the proppant placement result or final conductivity at a certain moment, making it difficult to reflect the dynamic changes between different stages during fracturing operations. Furthermore, existing models often fail to fully consider the close relationship between proppant placement and conductivity, making it easy for the predicted conductivity to deviate from the actual placement state. Moreover, their applicability and generalization ability under complex multi-level branched fracture networks remain insufficient.

[0007] Therefore, it is necessary to propose a joint prediction method that enables dynamic prediction of the fracturing construction stage and reflects the intrinsic relationship between proppant placement and conductivity. This method can achieve rapid and accurate prediction of proppant placement morphology and conductivity distribution in complex multi-level branched fracture networks, providing a reliable basis for optimizing fracturing construction schemes, controlling construction stages, and making real-time adjustments on site.

[0008] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0009] The main objective of this invention is to provide a prediction method based on deep learning for the combined proppant placement and flow capacity of multi-level branch joints, aiming to solve the aforementioned technical problems in the prior art.

[0010] To achieve the above objectives, this invention provides a deep learning-based method for predicting the joint proppant placement and flow conductivity of multi-level branched joints, comprising: The parameters of the branch fracture, construction parameters and reservoir parameters are sampled and combined to generate multiple simulation schemes covering various working conditions. Each simulation scheme includes a combination of the parameters of the branch fracture, construction parameters and reservoir parameters. The generated simulation scheme is input into a pre-constructed numerical model of proppant transport and placement covering multi-level branch seams. The simulation yields the two-dimensional matrix of proppant placement and the two-dimensional matrix of flow conduction capacity at each time step, forming a time series training dataset. A parameterized multi-task U-shaped convolutional neural network is constructed, comprising a one-dimensional vector input layer, a two-dimensional matrix input layer, a CNN shared feature extraction module, a time series module, and a dual UNet output module. The one-dimensional vector input layer is used to input branch fracture parameter encoding, construction parameters, and reservoir parameters; the two-dimensional matrix input layer is used to input the proppant placement matrix and conductivity matrix of the previous time step. The CNN shared feature extraction module performs convolution operations on the two-dimensional matrix input to extract the local spatial distribution features of proppant placement and conductivity on a regular grid in the fracture plane. The one-dimensional input vector... By mapping and expanding the fully connected layer into a two-dimensional parametric feature map with the same spatial size as the convolutional feature map, the local spatial features are fused with the two-dimensional parametric feature map, and a joint feature map sequence is obtained through further convolution. The joint feature map sequence is further input into the time series module, and the dynamic feature map output by the time series module is further input into the dual U-Net output module for decoding. U-Net1 is used to predict the two-dimensional matrix of proppant placement corresponding to each time step, and U-Net2 is used to predict the two-dimensional matrix of flowability corresponding to each time step. U-Net2 fuses the proppant placement matrix output by U-Net1 in the decoding stage. The parameterized multi-task U-shaped convolutional neural network is trained using the formed time-series training dataset; The parameters of the branch fractures to be predicted, the construction parameters, the reservoir parameters, and the initial state are input into a trained parameterized multi-task U-shaped convolutional neural network, and the proppant placement matrix and the conductivity matrix at each time step are obtained recursively.

[0011] Preferably, in the deep learning-based prediction method for the joint proppant placement and flowability of multi-level branch joints, the total loss function of the parameterized multi-task U-shaped convolutional neural network includes proppant placement loss, flowability loss, stage evolution consistency loss, placement-flowability coupling consistency loss, and physical constraint loss. Specifically, the proppant placement loss constrains the error between the predicted proppant placement matrix and the actual proppant placement matrix; the flowability loss constrains the error between the predicted flowability matrix and the actual flowability matrix; the stage evolution consistency loss constrains the trend of predicted changes between adjacent construction stages to remain consistent with the actual trend; the placement-flowability coupling consistency loss constrains the predicted flowability result to remain consistent with the empirical flowability result derived from the predicted proppant placement result; and the physical constraint loss ensures that the prediction result conforms to the laws of fluid flow and particle transport.

[0012] Preferably, in the deep learning-based prediction method for the combined proppant placement and flow capacity of multi-level branched joints, the formula for calculating the total loss function is: ; Among them, L S For proppant placement loss; L K This results in a loss of diversion capacity. L T This represents the consistency loss during stage evolution. L C To mitigate the loss of consistency in the flow-guiding coupling; L phy Loss due to physical constraints; , , , , These are the weighting coefficients for proppant placement loss, flow capacity loss, stage evolution consistency loss, placement-flow coupling consistency loss, and physical constraint loss, respectively.

[0013] Preferably, in the deep learning-based prediction method for the joint proppant placement and flow capacity of multi-level branched joints, the function of the proppant placement loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; S t The actual proppant placement matrix at time step t.

[0014] Preferably, in the deep learning-based prediction method for the joint prediction of multi-level branched joint proppant placement and flow capacity, the function of flow capacity loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; K t Let be the actual flow control capability matrix at time step t.

[0015] Preferably, in the deep learning-based prediction method for the joint prediction of multi-level branched joint proppant placement and flow conduction capacity, the function of the stage evolution consistency loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; The proppant placement matrix predicted at time step t-1; S t The actual proppant placement matrix at time step t; S t-1 The actual proppant placement matrix at time step t-1; Let be the flow capacity matrix predicted at time step t; This is the predicted flow capacity matrix at time step t-1; K t The actual flow capability matrix at time step t. K t-1 This is the actual flow capacity matrix at time step t-1.

[0016] Preferably, in the deep learning-based prediction method for the joint prediction of multi-level branched joint proppant placement and flow conduction capacity, the function of the placement-flow conduction coupling consistency loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; ; in, For the t-th time step The flow guidance capability matrix of each grid cell; Indicates the t-th time step. Predicted proppant placement value for each grid cell; Indicates the first Effective crack width of each grid cell; Indicates the first Normal closing stress of each grid cell; a, b, c, and d are coefficients.

[0017] Preferably, in the deep learning-based prediction method for the joint prediction of multi-level branched joint proppant placement and flow conduction capacity, the function of the physical constraint loss is: ; in, These are the residual terms of the fluid control equations; This refers to the residual term of the particle motion equation; , These are the weighting coefficients; ; ; For fluid velocity; p represents pressure; For fluid density; For fluid viscosity; The velocity of the nth particle; , , , These represent drag force, buoyancy, gravity, and contact force, respectively. f is the force per unit volume of fluid; m is the mass of a single proppant particle; N is the number of sampling points; t is a continuous time variable.

[0018] Preferably, in the deep learning-based prediction method for the combined proppant placement and diversion capacity of multi-level branched fractures, the numerical model for proppant transport and placement covering multi-level branched fractures includes a main fracture and first-, second-, and third-level branched fractures. The length, angle, and connection position of each branched fracture with the main fracture or the superior branched fracture are all adjustable parameters. For branched fractures that do not exist in the simulation scheme, the corresponding positions in the model are left blank or given zero-concentration boundary conditions.

[0019] Preferably, in the deep learning-based prediction method for the joint proppant placement and conductivity of multi-level branched joints, the simulation to obtain the two-dimensional matrix of proppant placement and the two-dimensional matrix of conductivity at each time step specifically includes: The proppant placement results and conductivity distribution results output from the numerical simulation are mapped onto a unified two-dimensional planar regular grid to form a two-dimensional proppant placement matrix and a two-dimensional conductivity matrix. Each matrix element corresponds to a fixed spatial location in the crack planar regular grid, and its value represents the proppant concentration or conductivity value at that location. The proppant placement matrix and conductivity matrix at each time step are normalized and arranged into a time series tensor in chronological order.

[0020] Preferably, in the deep learning-based prediction method for the joint proppant placement and conductivity of multi-level branched joints, the recursive acquisition of the proppant placement matrix and conductivity matrix at each time step includes: Using the zero matrix or the initial stage state matrix as initial condition input, the proppant placement matrix and flow capacity matrix of the first time step are obtained; The prediction results of the previous time step are used as the conditional input for the next time step, and the proppant placement matrix and flow capacity matrix for each subsequent time step are obtained by recursion step by step.

[0021] The present invention has at least the following beneficial effects: This invention first systematically samples and combines parameters of branch fractures (number of branches, length, angle, connection location), construction parameters (pump speed, pumping stage, proppant concentration and particle size), and reservoir parameters (Young's modulus, geostress, wall roughness, etc.) to generate a simulation scheme covering typical working conditions and extreme cases. This scheme is then input into a pre-constructed parameterized multi-level branch fracture proppant transport numerical model to efficiently simulate the two-dimensional matrix of proppant placement and the two-dimensional matrix of conduction capacity at each time step, forming a large-scale training dataset containing information on time evolution and multi-physics coupling. Secondly, a parameterized multi-task U-shaped convolutional neural network is constructed. This network receives the branch fracture code and construction / reservoir parameters through a one-dimensional vector input layer, and receives the proppant placement and diversion capacity matrices from the previous time step through a two-dimensional matrix input layer. It uses a CNN shared feature extraction module to capture the local spatial distribution features on the fracture plane and fuses the global parameter feature map. Then, a time series module (such as ConvLSTM) learns the dynamic evolution law between construction stages. Finally, the dual UNet output modules predict the proppant placement matrix and diversion capacity matrix of the current time step, respectively. Moreover, the diversion capacity prediction branch explicitly fuses the proppant placement information to achieve consistent joint prediction of two strongly correlated tasks. Furthermore, a multi-objective total loss function, including proppant placement loss, conductivity loss, stage evolution consistency loss, placement-conductivity coupling consistency loss, and physical constraint loss, is used to train the network. The first two ensure numerical accuracy, the stage evolution consistency loss constrains the physical trend of time-varying changes, the coupling consistency loss enforces the empirical relationship between conductivity and placement results, and the physical constraint loss embeds the conservation and reasonable laws of fluid flow and particle transport. This allows the network to not only fit the data in prediction but also conform to the underlying physical mechanisms, significantly improving prediction accuracy and generalization ability. Finally, in practical applications, only the branch fracture parameters, construction parameters, reservoir parameters, and initial state to be predicted need to be input into the trained network. Through a recursive approach (using the prediction results of the previous time step as the input conditions for the next time step), the proppant placement and conductivity distribution matrix for the entire fracturing operation process is rapidly generated. The computation time is far lower than traditional numerical simulations, meeting the needs of real-time on-site control. This enables rapid and accurate joint prediction of proppant placement morphology and conductivity in complex fracture networks with multi-level branching fractures, providing a reliable quantitative basis for offline optimization of fracturing construction schemes, online control during construction, and real-time parameter adjustment on site.

[0022] Furthermore, this invention enables joint and dynamic prediction of proppant placement morphology and conductivity distribution in complex multi-level branched fracture networks. Unlike existing methods that primarily perform static predictions targeting a single moment or final result, this invention treats key construction stages in the fracturing process as time-series nodes. Through a time-series module, it learns the migration, deposition, and redistribution patterns of proppant across different construction stages, thereby achieving stage-by-stage prediction of proppant placement and conductivity throughout the entire construction process. This method better aligns with the actual phased implementation of fracturing injection procedures, providing a direct basis for parameter adjustment and real-time optimization during on-site construction.

[0023] Furthermore, the flow capacity prediction is designed as a conditional generation process constrained by the proppant placement results. Specifically, the network first predicts the proppant placement matrix for each construction stage, and then the flow capacity prediction branch generates the corresponding flow capacity matrix based on the fusion of proppant placement results and time-series dynamic characteristics. This makes the flow capacity prediction no longer a separate output independent of proppant placement, but directly constrained by the spatial distribution state of the proppant, ultimately improving the physical consistency and engineering reliability of the flow capacity prediction results.

[0024] Furthermore, this invention also incorporates proppant placement error, flow-guiding capacity error, construction stage evolution consistency constraints, placement-flow-guiding coupling consistency constraints, and physical constraints into the network optimization process by constructing a joint loss function. This not only improves the accuracy of prediction results for individual construction stages but also enhances the model's ability to represent dynamic trends between different construction stages, making the prediction results more consistent with the proppant migration law, flow-guiding capacity formation law, and the real evolution characteristics of the on-site construction process.

[0025] Furthermore, this invention constructs multi-stage time series training samples based on numerical simulation results. After the network training is completed, only the branch fracture parameters, construction parameters, reservoir parameters, and initial conditions need to be input to recursively obtain the proppant placement matrix and conductivity matrix for each construction stage. Compared with traditional numerical simulations that require repeated calculations for each fracture combination and construction condition, this invention significantly shortens the prediction time, reduces computational costs, and has high field application value. It can be used for fracturing construction optimization, stage control, and real-time decision support. Attached Figure Description

[0026] Figure 1 The flowchart shows the prediction method for the combined proppant placement and flow capacity of multi-level branch joints based on deep learning provided by the present invention.

[0027] Figure 2 This is a comparison of the numerical simulation results and prediction results of the main seam flow capacity distribution cloud map according to one embodiment of the present invention. Figure 2(a) shows the numerical simulation results. Figure 2 (b) shows the prediction results. Figure 3 This is a schematic diagram of the parameterized multi-task U-shaped convolutional neural network provided by the present invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] In this embodiment of the invention, the term "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0032] In this embodiment of the invention, the term "multiple" refers to two or more, and other quantifiers are similar.

[0033] In this invention, unless otherwise stated, directional terms such as "upper," "lower," "top," and "bottom" are generally used in relation to the direction shown in the accompanying drawings, or in relation to the vertical, perpendicular, or gravitational direction of the component itself; similarly, for ease of understanding and description, "inner" and "outer" refer to the inner and outer contours of each component itself, but the above directional terms are not intended to limit this invention.

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0035] Figure 1 The flowchart illustrates a first embodiment of the deep learning-based prediction method for the joint proppant placement and flowability of multi-level branch joints provided by the present invention. This deep learning-based prediction method for the joint proppant placement and flowability of multi-level branch joints includes steps S1000 to S4000.

[0036] Step S1000 samples and combines the branch fracture parameters, construction parameters, and reservoir parameters to generate multiple simulation schemes covering various working conditions. Each simulation scheme includes a combination of branch fracture parameters, construction parameters, and reservoir parameters.

[0037] Specifically, step S1000 includes determining a reasonable range of values ​​for branch fracture parameters (including the number of branches, the length of each branch fracture, the angle between the branch fracture and the main fracture or the superior branch fracture, and the connection position of the branch fracture) based on on-site geological exploration data, historical construction records, and fracture development patterns. A spatial uniform sampling method is then used to independently sample these parameters within this range to cover common actual distributions and potential extreme situations. Simultaneously, based on actual well section construction records and geological exploration data, the minimum, maximum, and typical values ​​of construction parameters (such as pump speed, number of pumping stages, proppant concentration, proppant particle size distribution, etc.) and reservoir parameters (such as Young's modulus, Poisson's ratio, geostress, fracture wall roughness, etc.) are statistically analyzed. A parameter value range covering typical working conditions and boundary conditions is designed, and multiple combinations of construction parameters and reservoir parameters are generated using appropriate sampling methods (such as Latin hypercube sampling or random sampling). Finally, the sampled combinations of branch fracture parameters are randomly matched with the combinations of construction and reservoir parameters to form a sufficient number of simulation schemes with diverse characteristics, which serve as input for subsequent numerical simulations.

[0038] More specifically, the sampling methods include, but are not limited to, full permutation sampling, Latin hypercube sampling, random sampling, and stratified sampling. Preferably, the present invention employs the Latin hypercube sampling method to generate multiple sets of construction parameters and reservoir parameter combinations within a preset parameter range, so as to ensure that each parameter is evenly distributed within its value range, and to take into account the coverage of training data while controlling the number of samples.

[0039] Step S2000 inputs the generated simulation scheme into the pre-constructed numerical model of proppant transport and placement covering multi-level branch seams, and simulates to obtain the two-dimensional matrix of proppant placement and the two-dimensional matrix of flow conduction capacity at each time step, forming a time series training dataset.

[0040] The numerical model for proppant transport and placement covering multi-level branch cracks includes a main crack and first-, second-, and third-level branch cracks. The length, angle, and connection position of each branch crack to the main crack or the upper-level branch crack are all adjustable parameters. For branch cracks that do not exist in the simulation scheme, the corresponding positions in the model are left blank or given zero-concentration boundary conditions.

[0041] The simulation yields two-dimensional proppant placement matrices and two-dimensional conductivity matrices for each time step. Specifically, this involves mapping the proppant placement results and conductivity distribution results from the numerical simulation onto a unified two-dimensional planar regular grid to form two-dimensional proppant placement matrices and two-dimensional conductivity matrices. Each matrix element corresponds to a fixed spatial location in the crack planar regular grid, and its value represents the proppant concentration or conductivity value at that location. The two-dimensional proppant placement matrices and two-dimensional conductivity matrices for each time step are normalized and arranged into a time series tensor in chronological order.

[0042] More specifically, all simulation schemes generated in step S1000 are input one by one into a pre-constructed parametric fracture numerical model. This model can dynamically generate the main fracture and first to third-order branch fractures, and automatically configure the fracture network according to the geometric parameters such as branch fracture length, angle, and connection position given in each scheme. For branch fractures that do not exist in the scheme, the model automatically leaves them blank or assigns zero-concentration boundary conditions to maintain the consistency of the model structure. Subsequently, based on fluid dynamics and particle transport theory, the model numerically simulates the entire process of fracturing fluid carrying proppant in the main fracture and each branch fracture under given construction parameters (such as pump speed, pumping program, proppant concentration and particle size distribution) and reservoir parameters (such as Young's modulus, Poisson's ratio, in-situ stress, and fracture wall roughness). The model outputs the proppant placement results and conductivity distribution results for each time step according to the construction stage. To further adapt to the input format of deep learning networks, the simulation results were mapped onto a unified two-dimensional planar regular grid, resulting in a two-dimensional matrix of proppant placement and a two-dimensional matrix of conductivity. Each matrix element corresponds to the proppant concentration or conductivity value at a fixed location within the fracture planar regular grid. After normalizing the matrices for all time steps, they were organized into a time series tensor in chronological order. Using the placement matrix and conductivity matrix of the previous time step as input features and the corresponding matrix of the current or next time step as supervision labels, this tensor, combined with the branch fracture parameters, construction parameters, and reservoir parameters of each scheme, constitutes a complete time series training dataset.

[0043] A numerical model for proppant migration and placement is constructed. This model is a parametric fracture model used to simulate the proppant migration, deposition, and placement processes in the main fracture and its multi-level branch fractures. The model includes the main fracture and first-, second-, and third-level branch fractures. The length, angle, and connection position of each branch fracture to the main fracture or superior branch fracture can be configured as adjustable parameters, thereby enabling the dynamic generation of different fracture combinations.

[0044] For branch cracks that are not present in the current simulation scheme, blank spaces can be left at the corresponding locations in the model or zero-concentration boundary conditions can be applied to maintain the consistency of the model structure. This parametric numerical model can cover different levels of branch cracks and their common geometric variations within a single numerical simulation framework, giving the model strong versatility and flexibility, and providing a reliable foundation for subsequent proppant migration and placement simulations.

[0045] More specifically, the simulation scheme is input into the above-mentioned proppant migration and placement numerical model. According to the branch joint combination, the non-existent crack locations are set to empty or zero concentration. The migration, deposition and placement process of proppant in the main crack and each branch crack is simulated, and the proppant placement results and conductivity distribution results corresponding to each construction stage are output, wherein each construction stage corresponds to a time step.

[0046] The numerical simulation results are preferably exported in the form of numerical field data and mapped onto a uniform two-dimensional planar regular grid, thereby forming a two-dimensional matrix of proppant placement and a two-dimensional matrix of conductivity. Each matrix element corresponds to a fixed spatial location within the fracture planar regular grid, and its value represents the proppant concentration or conductivity value at that location. Preferably, the conductivity value of the corresponding grid cell can be calculated based on the proppant concentration within each grid cell, combined with fracture geometry parameters, closure conditions, and a preset conductivity calculation model or empirical relationship.

[0047] After obtaining the two-dimensional proppant placement matrix and the two-dimensional conductivity matrix at each time step, they are normalized and arranged into time series tensors in chronological order to form complete time series matrix data. During training, the proppant placement matrix and conductivity matrix of the previous time step are used as conditional inputs, and the proppant placement matrix and conductivity matrix corresponding to the current or next time step are used as supervision labels. Combined with branch joint parameters, construction parameters, and reservoir parameters, a complete training dataset is constructed.

[0048] Finally, the training dataset is divided into training, validation, and test sets according to a certain ratio, and it is ensured that the number of fracture levels, branch parameters, construction parameters, and reservoir parameters of each branch are covered in each dataset. This avoids the situation where a certain type of fracture combination or construction condition only appears in the training set, thereby ensuring the sufficiency of model training and the generalization ability of prediction results.

[0049] It should be noted that a parameterized numerical model of proppant migration and placement covering multi-level branch fractures was constructed. This model, based on a CFD-DEM coupling method and combining the numerical simulation capabilities of EDEM and Fluent software, is used to characterize the proppant migration, deposition, and placement processes within the fracture network. The model includes the main fracture and first-, second-, and third-order branch fractures. The length, angle, and connection position of each branch fracture to the main fracture or higher-level branch fractures can be configured as adjustable parameters, thereby enabling the dynamic generation of different branch combinations.

[0050] Specifically, the script interface or parametric geometry generation function of CFD-DEM can be used to dynamically adjust the crack geometry and injection boundary conditions according to different combinations of branch parameters, thereby realizing the simulation of multiple branch combinations under the same numerical model framework, without having to build an independent model for each branch combination, thus improving simulation efficiency and expanding data coverage.

[0051] To ensure the integrity and simulability of the model structure, all potential branch locations are pre-defined in the model. For branch cracks that do not exist in the current simulation scheme, the corresponding locations are left blank or given zero-concentration boundary conditions to reflect the actual possible branch absence. Through the above parametric modeling method, the constructed numerical model can simulate the proppant migration, deposition, and placement behavior under crack networks of different complexities, and provide a unified and reliable physical basis for the generation of subsequent training data.

[0052] It should be noted that the aforementioned parametric numerical model is mainly used to generate high-precision numerical simulation results under different crack combinations; while the branch crack parametric encoding described later is used to input the same crack geometric information into the deep learning network in the form of digital vectors. The two correspond in their parameter meanings, but functionally they are used for training label generation in the numerical simulation stage and input representation in the network training stage, respectively.

[0053] Each simulation scheme is input into the parameterized proppant migration and placement numerical model. For branch fractures not present in each scheme, the corresponding positions in the model are left blank or zero-concentration boundary conditions are assigned to reflect the actual possible branch absence and ensure that the simulation results correspond one-to-one with the current fracture combination. After running the numerical model, the migration, deposition, and placement process of proppant in the main fracture and each branch fracture is simulated, and the corresponding proppant placement results and conductivity distribution results are output according to the construction injection stage. Each construction injection stage corresponds to a key time step, which is used to characterize the dynamic state of proppant injection and deposition at that stage. Preferably, the key construction injection stages include the pre-injection stage, the low sand ratio stage, the medium sand ratio stage, the high sand ratio stage, and the pump shutdown stage.

[0054] The proppant placement and conductivity results are preferably exported as numerical field data calculated using coupled EDEM and Fluent software. To generate input data suitable for deep learning network training, a unified two-dimensional coordinate system is first established along the plane of the main fracture extension direction and the branch fractures, and the numerical results of each time step are mapped onto this two-dimensional plane. Subsequently, the two-dimensional plane is discretized into a regular grid according to a preset spatial resolution, allowing different samples to be represented under a unified spatial size and resolution. For the proppant placement results, the proppant concentration value within each grid cell is extracted to form a two-dimensional proppant placement matrix. For the conductivity results, based on the proppant concentration within each grid cell, and combined with fracture geometric parameters, closure conditions, and a preset conductivity calculation model or empirical relationship, the conductivity value of the corresponding grid cell is obtained, thus forming a two-dimensional conductivity matrix. In the two-dimensional matrix, each matrix element corresponds to a fixed spatial position in the regular grid of the fracture plane, and its value represents the proppant concentration or conductivity at that position.

[0055] After obtaining the two-dimensional proppant placement matrix and conductivity matrix for each time step, these matrices are normalized and arranged chronologically into [time_steps, height, width] tensors, forming a complete time-series matrix data. Further, when constructing training samples, branch fracture parameters, construction parameters, and reservoir parameters are used as network input features, with the proppant placement matrix and conductivity matrix of the previous time step as conditional inputs. The proppant placement matrix and conductivity matrix corresponding to the current or next time step are used as supervision labels, thus constructing a complete time-series training dataset. In this way, the network learns the mapping relationship between the fracture state, branch fracture parameters, construction parameters, and reservoir parameters of the previous time step and the proppant placement and conductivity results of the current time step.

[0056] In one optional implementation, the numerical results can be further generated into color images for visualization or result verification; when the software output results are mainly saved in image form, the two-dimensional numerical matrix corresponding to the regular grid can also be recovered through image processing methods.

[0057] Finally, the training dataset is divided into training, validation, and test sets according to a certain ratio, and it is ensured that the number of branch fractures, branch parameters, construction parameters, and reservoir parameters are all covered in each dataset, thereby ensuring the sufficiency of network training and the generalization ability of prediction results.

[0058] Step S3000 constructs a parameterized multi-task U-shaped convolutional neural network, which includes a one-dimensional vector input layer, a two-dimensional matrix input layer, a CNN shared feature extraction module, a time series module, and a dual UNet output module. For example... Figure 3 As shown, the one-dimensional vector input layer is used to input the branch fracture parameter encoding, construction parameters, and reservoir parameters, while the two-dimensional matrix input layer is used to input the proppant placement matrix and conductivity matrix of the previous time step. The CNN shared feature extraction module performs convolution operations on the two-dimensional matrix input to extract the local spatial distribution features of proppant placement and conductivity on the regular grid of the fracture plane. The one-dimensional input vector is mapped and expanded into a two-dimensional parametric feature map with the same spatial size as the convolution feature map through a fully connected layer. The local spatial features are fused with the two-dimensional parametric feature map, and a joint feature map sequence is obtained through further convolution. The joint feature map sequence is further input into the time series module, and the dynamic feature map output by the time series module is further input into the dual U-Net output module for decoding. U-Net1 is used to predict the two-dimensional proppant placement matrix corresponding to each time step, and U-Net2 is used to predict the two-dimensional conductivity matrix corresponding to each time step. U-Net2 fuses the proppant placement matrix output by U-Net1 in the decoding stage.

[0059] A parameterized multi-task U-shaped convolutional neural network is constructed, which consists of a one-dimensional vector input layer, a two-dimensional matrix input layer, a CNN shared feature extraction module, a time series module, and a dual UNet output module. Specifically, the one-dimensional vector input layer receives encoded branch fracture parameters (the length, angle, and connection position of each branch fracture are vectorized and encoded, and the encoding of non-existent branch fractures is set to zero), construction parameters (such as pump rate, proppant concentration, particle size distribution, etc.), and reservoir parameters (such as Young's modulus, Poisson's ratio, in-situ stress, etc.). These one-dimensional features are then mapped and expanded into a two-dimensional parametric feature map with the same spatial size as the two-dimensional feature map through a fully connected layer to represent global control information. The two-dimensional matrix input layer receives the proppant placement two-dimensional matrix and the conductivity two-dimensional matrix from the previous time step as the spatial distribution conditions of the current fracture state. The CNN shared feature extraction module uses convolutional operations to perform hierarchical feature extraction on the two-dimensional input matrix, capturing the local spatial distribution patterns of proppant placement and flowability on a regular grid in the crack plane, and outputting a convolutional feature map. The two-dimensional parametric feature map and the convolutional feature map are then fused along the channel dimension and further convolved to obtain a joint feature map. This joint feature map is input sequentially into the time-series module (preferably ConvLSTM or Transformer) to capture the dynamic temporal dependencies between proppant placement and flowability as construction progresses. Finally, the dynamic feature map output from the time-series module is fed into two decoders in the dual UNet output module: the first UNet decoder (UNet1) predicts the two-dimensional proppant placement matrix at the current time step, and the second UNet decoder (UNet2) predicts the two-dimensional flowability matrix at the current time step. UNet2 explicitly fuses the intermediate layer features or the final proppant placement prediction result from UNet1 during the decoding stage, thus subjecting the flowability prediction to the physical constraints of the proppant placement results, achieving joint prediction and enhanced consistency between the two tasks.

[0060] Step S4000 uses the formed time-series training dataset to train the parameterized multi-task U-shaped convolutional neural network. The total loss function includes proppant placement loss, flowability loss, stage evolution consistency loss, placement-flow coupling consistency loss, and physical constraint loss. The proppant placement loss is used to constrain the error between the predicted proppant placement matrix and the actual proppant placement matrix. The flowability loss is used to constrain the error between the predicted flowability matrix and the actual flowability matrix. The stage evolution consistency loss is used to constrain the change trend of the predicted results between adjacent construction stages to be consistent with the actual change trend. The placement-flow coupling consistency loss is used to constrain the predicted flowability result to be consistent with the empirical flowability result derived from the predicted proppant placement result. The physical constraint loss is used to ensure that the predicted result conforms to the fluid flow law and particle transport law.

[0061] The formula for calculating the total loss function is as follows: ; Among them, L S For proppant placement loss; L K This results in a loss of diversion capacity. L T This represents the consistency loss during stage evolution. L C To mitigate the loss of consistency in the flow-guiding coupling; L phy Loss due to physical constraints; , , , , These are the weighting coefficients for proppant placement loss, flow capacity loss, stage evolution consistency loss, placement-flow coupling consistency loss, and physical constraint loss, respectively.

[0062] The function of the proppant placement loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; S t The actual proppant placement matrix at time step t.

[0063] The function of the loss of the flow guiding capacity is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; K t Let be the actual flow control capability matrix at time step t.

[0064] The stage evolution consistency loss is used to constrain the predicted trend of changes between adjacent construction stages to maintain consistency with the actual trend, thereby strengthening the time series module's learning of the dynamic evolution law of construction stages. The function of the stage evolution consistency loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; The proppant placement matrix predicted at time step t-1; S t The actual proppant placement matrix at time step t; S t-1 The actual proppant placement matrix at time step t-1; Let be the flow capacity matrix predicted at time step t; This is the predicted flow capacity matrix at time step t-1; K t The actual flow capability matrix at time step t. K t-1 This is the actual flow capacity matrix at time step t-1.

[0065] The proppant placement coupling consistency loss is used to ensure that the predicted proppant placement results are consistent with the empirical proppant placement results derived from the predicted proppant placement results, thereby strengthening the constraint relationship between the proppant placement morphology and proppant placement coupling consistency. The function of the proppant placement coupling consistency loss is: ; Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; ; in, For the t-th time step The flow guidance capability matrix of each grid cell; Indicates the t-th time step. Predicted proppant placement value for each grid cell; Indicates the first Effective crack width of each grid cell; Indicates the first Normal closing stress of each grid cell; a, b, c, and d are coefficients.

[0066] The function of the physical constraint loss is: ; in, These are the residual terms of the fluid control equations; This refers to the residual term of the particle motion equation; , These are the weighting coefficients; ; ; For fluid velocity; p represents pressure; For fluid density; For fluid viscosity; The velocity of the nth particle; , , , These represent drag force, buoyancy, gravity, and contact force, respectively. f is the force per unit volume of fluid; m is the mass of a single proppant particle; N is the number of sampling points; t is a continuous time variable.

[0067] Each branch fracture within the fracture system is parametrically encoded, including its length, the angle between the branch fracture and the main fracture or superior branch fracture, and the connection point between the branch fracture and the main fracture or superior branch fracture. If a branch fracture is absent in a training sample, its corresponding code is set to zero or assigned a special label so that the network can recognize it as an empty branch. The encoded branch fracture parameters, along with the construction parameters and reservoir parameters, constitute the one-dimensional input features of the network.

[0068] The constructed deep learning network includes an input module, a CNN shared feature extraction module, a time series module, and a dual-U-Net multi-task output module. The input module comprises two parts: a one-dimensional vector input layer for inputting fracture coding, construction parameters, and reservoir parameters; and a two-dimensional matrix input layer for inputting the proppant placement matrix and conductivity matrix from the previous time step, characterizing the current fracture state and providing conditional information for prediction in subsequent time steps. Both the proppant placement matrix and conductivity matrix are physical matrices formed by mapping CFD-DEM numerical results to a unified two-dimensional planar regular grid, where each matrix element corresponds to the proppant concentration or conductivity value at a fixed spatial location in the fracture plane.

[0069] During feature extraction, the CNN shared feature extraction module performs convolution operations on the two-dimensional matrix input to extract the local spatial distribution features of proppant placement and conductivity on a regular grid of fracture planes. Simultaneously, the one-dimensional input vector is mapped and expanded through a fully connected layer into a two-dimensional parametric feature map with the same spatial size as the convolutional feature map, used to characterize global control information such as branch fracture geometry parameters, construction parameters, and reservoir parameters. Subsequently, the local spatial features and the two-dimensional parametric feature map are fused, and a joint feature map is obtained through further convolution. The local spatial distribution features include variations in proppant concentration within the main fracture and branch fracture regions, changes in numerical gradients near branch connection points, continuous or attenuated distributions of conductivity within local channels, and spatial response features of the intersection regions of the main fracture and multi-level branch fractures. At the same time, the one-dimensional input vector is mapped and expanded through a fully connected layer into a two-dimensional parametric feature map with the same spatial size as the convolutional feature map, used to supplement the network with global control information such as fracture geometry parameters, construction conditions, and reservoir conditions. Subsequently, the local spatial features are fused with the two-dimensional parametric feature map, and a joint feature map is obtained through further convolution, so that each spatial location simultaneously contains local spatial distribution information within the fracture plane as well as global information such as branch fracture geometric parameters, construction parameters, and reservoir parameters.

[0070] The joint feature map sequence is further input into a time series module to capture the dynamic patterns of proppant placement and diffusivity evolution during construction stages. Preferably, the time series module is a ConvLSTM, but a Transformer module can also be used. In this embodiment, a ConvLSTM time series module is used. The dynamic feature map output by the time series module is further input into a dual U-Net decoder, where U-Net1 is used to predict the two-dimensional matrix of proppant placement at each time step, and U-Net2 is used to predict the two-dimensional matrix of diffusivity at each time step. U-Net2 fuses the proppant placement matrix output by U-Net1 during the decoding stage, making the diffusivity prediction explicitly constrained by the proppant placement results, thereby improving the physical consistency and prediction accuracy of the diffusivity prediction results and achieving joint prediction of proppant placement and diffusivity.

[0071] Unlike conventional static prediction models, the ConvLSTM time series module in this invention uses key construction stages in the fracturing process as time series nodes. These key construction stages preferably include the pre-injection stage, the low-spar ratio stage, the medium-spar ratio stage, the high-spar ratio stage, and the pump shutdown stage. Each stage corresponds to a time step, used to characterize the migration, deposition, and placement of proppant in the main fracture and branch fractures, as well as the formation and change of corresponding conductivity. The ConvLSTM time series module is used to learn the dynamic evolution between different construction stages, specifically including: the cumulative effect of the proppant placement state formed in the previous construction stage on the migration, deposition, and redistribution of proppant in the main fracture and branch fractures in the subsequent construction stage; the driving effect of changes in pump rate, proppant concentration, and pumping procedure on the proppant transport process from the main fracture to the branch fracture; and the staged influence of changes in proppant placement state on the formation, enhancement, or attenuation of conductivity. Through the aforementioned time series modeling, the network can learn the correspondence between the fracture state, branch fracture parameters, construction parameters, and reservoir parameters of the previous construction stage and the proppant placement and conductivity results of the current construction stage, thereby achieving dynamic recursive prediction of the entire fracturing construction process.

[0072] The time-series dynamic feature maps output by ConvLSTM are further input into a dual U-Net decoder. U-Net1 is used to predict the two-dimensional proppant placement matrix corresponding to each construction stage. Its structure includes three parts: an encoder, skip connections, and a decoder. The encoder consists of multiple convolutional and pooling layers to extract multi-scale spatial features. Skip connections connect the outputs of each encoder layer to the corresponding layers of the decoder to preserve high-resolution local information. The decoder recovers the matrix size through upsampling and convolution, and fuses the skip connection features to output the proppant placement prediction matrix corresponding to each construction stage.

[0073] The time-series dynamic feature maps output by ConvLSTM are further input into a dual U-Net decoder. U-Net1 is used to predict the two-dimensional proppant placement matrix corresponding to each construction stage. Its structure consists of three parts: an encoder, skip connections, and a decoder. The encoder comprises multiple convolutional and pooling layers to extract multi-scale spatial features. Skip connections connect the outputs of each encoder layer to the corresponding layers of the decoder to preserve high-resolution local information. The decoder recovers the matrix size through upsampling and convolution, and fuses the skip connection features to output the proppant placement prediction matrix corresponding to each construction stage.

[0074] U-Net2 is used to predict the two-dimensional conductivity matrix corresponding to each construction stage. However, its prediction process is not independent of the proppant placement prediction; rather, it is a conditional generation process based on the dynamic constraints of the proppant placement results. Specifically, U-Net2 fuses the time-series dynamic feature map output by ConvLSTM and the proppant placement matrix output by U-Net1 during the decoding stage. This makes the conductivity prediction not only dependent on spatiotemporal feature information but also on the spatial distribution of proppant in the main and branch cracks at each construction stage. Since the formation of crack conductivity is essentially influenced by the proppant concentration distribution, crack geometry, and closure conditions, incorporating the proppant placement results into the conductivity prediction branch allows the conductivity prediction process to be directly constrained by the proppant placement state. This avoids the conductivity prediction results deviating from the actual placement state, improving the physical consistency and engineering reliability of the conductivity prediction. Preferably, the two-dimensional conductivity matrix generated by U-Net2 is not only subject to error constraints with the actual conductivity label, but also to consistency constraints with the empirical conductivity results calculated based on the proppant placement results, crack geometry parameters, and closure conditions output by U-Net1. This allows for a better reflection of the actual impact of proppant placement on conductivity formation in terms of both network structure and loss function.

[0075] More specifically, the parameterized multi-task U-shaped convolutional neural network is trained using a training set. Forward propagation yields the predicted proppant placement matrix and conductivity matrix, and backpropagation, combined with a total loss function, updates the network weights. The total loss function includes proppant placement loss, conductivity loss, stage evolution consistency loss, placement-conduction coupling consistency loss, and physical constraint loss. This allows the network to learn the variations in proppant placement and conductivity with construction parameters, reservoir conditions, and branch fracture parameters, while maintaining the physical reasonableness of the predicted results.

[0076] Among them, proppant placement loss is used to constrain the error between the predicted proppant placement matrix and the actual matrix; flowability loss is used to constrain the error between the predicted flowability matrix and the actual matrix; stage evolution consistency loss is used to constrain the changing trend of the prediction results between adjacent construction stages; placement-flow coupling consistency loss is used to constrain the coupling relationship between the flowability prediction results and the proppant placement results; and physical constraint loss is used to ensure that the prediction results conform to the fluid flow law and particle transport law.

[0077] During training, the validation set is used to monitor model performance, prevent overfitting, and adjust the learning rate, optimize hyperparameters, and control early stopping based on changes in the validation set loss. The validation set contains different parameter combinations than the training set, used to evaluate the model's generalization ability and assist in selecting the optimal model weights during training. The test set is used for final performance evaluation after training, to verify the model's prediction accuracy and generalization ability under unseen parameter combinations. Output metrics include mean squared error (MSE) and mean absolute error (MAE).

[0078] After completing the network structure construction, the parameterized multi-task U-shaped convolutional neural network was trained using the time series training dataset obtained in step 3. During the training phase, branch joint parameters, construction parameters, and reservoir parameters were used as input features, the proppant placement matrix and conductivity matrix of the previous time step were used as conditional inputs, and the proppant placement matrix and conductivity matrix of the corresponding time step were used as supervision labels. The network prediction results were obtained through forward propagation.

[0079] Step S5000: Input the branch fracture parameters, construction parameters, reservoir parameters and initial state to be predicted into the trained parameterized multi-task U-shaped convolutional neural network, and recursively obtain the proppant placement matrix and conductivity matrix for each time step.

[0080] The recursive method for obtaining the proppant placement matrix and conductivity matrix at each time step includes: Using the zero matrix or the initial stage state matrix as initial condition input, the proppant placement matrix and flow capacity matrix of the first time step are obtained; The prediction results of the previous time step are used as the conditional input for the next time step, and the proppant placement matrix and flow capacity matrix for each subsequent time step are obtained by recursion step by step.

[0081] During the prediction phase, construction parameters, reservoir parameters, and branch fracture codes are input into the network, and the zero matrix or initial stage state matrix is ​​used as the initial condition input to obtain the proppant placement matrix and conductivity matrix for the first time step. Subsequently, the prediction results of the previous time step are used as the condition input for the next time step, and the prediction results of proppant placement and conductivity for subsequent time steps are obtained recursively, thereby realizing dynamic prediction of the entire fracturing construction process.

[0082] Example To verify the effectiveness of the proposed method for predicting the placement morphology and conductivity of intra-surgery proppant based on multi-task deep learning, the following embodiments are designed.

[0083] Table 1 Relevant Reservoir and Fracturing Operation Parameters Table 2 Taking the fractured section of a horizontal well in a shale reservoir as an example, this paper describes the method for joint prediction of multi-level branch fracture proppant placement and conductivity proposed in this invention.

[0084] First, reservoir parameters of the block containing the target well section are obtained, including Young's modulus, Poisson's ratio, maximum in-situ stress, and fracture wall roughness. Simultaneously, fracturing parameters are acquired, including pump rate, number of injection stages, proppant concentration, and proppant particle size. Further, branch fracture parameters are obtained, including the lengths of first-, second-, and third-order branch fractures, as well as the angles and connection points between each branch fracture and the main fracture or superior branch fracture, to characterize the geometric features of the complex fracture network.

[0085] Based on the aforementioned reservoir parameters, construction parameters, and branch fracture parameter ranges, 1000 sets of simulated operating conditions were generated using the Latin hypercube sampling method. Each set of simulated operating conditions includes a set of branch fracture combinations, a set of construction parameter combinations, and a set of reservoir parameter combinations, covering the main operating conditions and some extreme conditions that may occur in this well section. Subsequently, the above 1000 sets of simulated operating conditions were input into a parameterized CFD-DEM coupled numerical model. The numerical model was jointly implemented using EDEM software and Fluent software to simulate the migration, deposition, and placement processes of proppant within the main fracture and multi-level branch fractures.

[0086] In this embodiment, the fracturing process is divided into five key stages: pre-injection, low-sand-ratio, medium-sand-ratio, high-sand-ratio, and pump shutdown. Each stage corresponds to a time step. For each simulated condition, the proppant placement and conductivity results for each of the five time steps are output. The proppant placement results are represented by the proppant concentration values ​​on a regular grid in the fracture plane; the conductivity results are obtained based on the proppant concentration within each grid cell, combined with fracture geometry, closure conditions, and a preset conductivity calculation relationship. To ensure comparability between different samples, the fracture plane is uniformly discretized into a 100×50 regular grid in this embodiment. Therefore, each time step corresponds to a 100×50 proppant placement matrix and a 100×50 conductivity matrix.

[0087] During data processing, the proppant placement and conductivity results at each time step are mapped onto a unified two-dimensional planar grid, forming two-dimensional matrices for proppant placement and conductivity, respectively, and then normalized. Since each simulation case contains 5 time steps, each case can be organized into time-series tensor data of size [5, 100, 50]. When constructing training samples, the proppant placement and conductivity matrices of the previous time step are used as input conditions, and the proppant placement and conductivity matrices corresponding to the current time step are used as supervision labels, thus forming time-series training samples of "stage t input - stage t+1 output". Therefore, each simulation case containing 5 time steps can form 4 sets of adjacent stage training samples, resulting in 4000 sets of time-series training samples from 1000 simulation cases. Subsequently, the dataset was divided according to the simulated working condition number, with 700 working conditions used as the training set, 150 working conditions used as the validation set, and 150 working conditions used as the test set, in order to avoid samples from different time steps under the same working condition appearing in both the training set and the test set at the same time.

[0088] In terms of model construction, this embodiment employs a parameterized multi-task U-shaped convolutional neural network. The network input consists of two parts: a one-dimensional parameter vector containing branch fracture codes, construction parameters, and reservoir parameters; and a two-dimensional matrix input containing the proppant placement matrix and conductivity matrix from the previous time step. The network first extracts local spatial distribution features on the regular grid of the fracture plane using a CNN shared feature extraction module. Simultaneously, the one-dimensional parameter vector is mapped to a two-dimensional parameter feature map via a fully connected layer and fused with the local spatial features. The fused joint feature map sequence is input into a ConvLSTM time series module to learn the dynamic patterns of proppant migration, deposition, and redistribution between different construction stages. Subsequently, the dynamic feature map is input into a dual U-Net decoder, where U-Net1 outputs the proppant placement prediction matrix for each time step, and U-Net2 outputs the conductivity prediction matrix for each time step. The output of U-Net1 is fused during the decoding stage, ensuring that the conductivity prediction is constrained by the proppant placement state.

[0089] During training, this embodiment uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 16, a gradient clipping threshold of 5, and a maximum training epoch of 500.

[0090] The total loss function comprises proppant placement loss, conductance capacity loss, stage evolution consistency loss, placement-conductance coupling consistency loss, and physical constraint loss. The total loss function can be expressed as: In one embodiment, the weights of each loss term can be initially set to α=1.0, β=1.0, γ=0.5, η=0.5, and ξ=0.1, and adjusted during training based on changes in the validation set loss.

[0091] During training, the validation set loss is used as the basis for model selection; when the validation set loss does not decrease for 20 consecutive rounds, training is stopped and the optimal model parameters are saved.

[0092] After training, test set samples are input into the network for prediction. In the prediction phase, branch fracture parameters, construction parameters, and reservoir parameters are first input, with a zero matrix as the initial condition input, to obtain the proppant placement matrix and conductivity matrix for the first time step. Subsequently, the prediction result of this time step is used as the condition input for the next time step, recursively obtaining the prediction results for subsequent time steps. Finally, the multi-time-step proppant placement distribution and conductivity distribution of the well section throughout the entire construction process can be obtained, and the corresponding two-dimensional visualization cloud map is further generated. Test results show that the model in this embodiment has good prediction effects on proppant placement and conductivity, with a determination coefficient R² exceeding 0.90, which can meet the application requirements for dynamic prediction and on-site auxiliary decision-making in fracturing construction under complex multi-level branch fracture networks.

[0093] This specific embodiment completes the joint prediction process of proppant placement morphology and conductivity in complex multi-level branch fracture networks through the above steps. It realizes the rapid acquisition of multi-time step two-dimensional prediction results after inputting construction parameters, reservoir parameters, branch fracture parameters and initial conditions, which can be used for fracturing construction optimization and real-time adjustment.

[0094] Figure 2 This is a comparison of the numerical simulation results and prediction results of the main seam flow capacity distribution cloud map according to one embodiment of the present invention. Figure 2 (a) shows the numerical simulation results. Figure 2 (b) shows the prediction results.

[0095] from Figure 2 It can be seen that the prediction results of the multi-level branch joint proppant placement and flow capacity prediction method based on deep learning provided by this invention are accurate.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting the combined proppant placement and flow conductivity of multi-level branched joints based on deep learning, characterized in that, include: The parameters of the branch fracture, construction parameters and reservoir parameters are sampled and combined to generate multiple simulation schemes covering various working conditions. Each simulation scheme includes a combination of the parameters of the branch fracture, construction parameters and reservoir parameters. The generated simulation scheme is input into a pre-constructed numerical model of proppant transport and placement covering multi-level branch seams. The simulation yields the two-dimensional matrix of proppant placement and the two-dimensional matrix of flow conduction capacity at each time step, forming a time series training dataset. A parameterized multi-task U-shaped convolutional neural network is constructed, which includes a one-dimensional vector input layer, a two-dimensional matrix input layer, a CNN shared feature extraction module, a time series module, and a dual UNet output module. One-dimensional vector input layer is used to input branch fracture parameter encoding, construction parameters and reservoir parameters, and two-dimensional matrix input layer is used to input the proppant placement matrix and conductivity matrix of the previous time step; the CNN shared feature extraction module performs convolution operation on the two-dimensional matrix input to extract the local spatial distribution features of proppant placement and conductivity on the regular grid of the fracture plane; The one-dimensional input vector is mapped and expanded into a two-dimensional parametric feature map with the same spatial size as the convolutional feature map through a fully connected layer. The local spatial features are fused with the two-dimensional parametric feature map, and a joint feature map sequence is obtained through further convolution. The joint feature map sequence is further input into the time series module. The dynamic feature map output by the time series module is further input into the dual U-Net output module for decoding. U-Net1 is used to predict the two-dimensional proppant placement matrix corresponding to each time step, and U-Net2 is used to predict the two-dimensional flow capacity matrix corresponding to each time step. U-Net2 fuses the proppant placement matrix output by U-Net1 in the decoding stage. The parameterized multi-task U-shaped convolutional neural network is trained using the formed time-series training dataset; The parameters of the branch fractures to be predicted, the construction parameters, the reservoir parameters, and the initial state are input into a trained parameterized multi-task U-shaped convolutional neural network, and the proppant placement matrix and the conductivity matrix at each time step are obtained recursively.

2. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 1, characterized in that, The total loss function of the parameterized multi-task U-shaped convolutional neural network includes proppant placement loss, conductivity loss, stage evolution consistency loss, placement-conduction coupling consistency loss, and physical constraint loss. Specifically, the proppant placement loss constrains the error between the predicted proppant placement matrix and the actual proppant placement matrix; the conductivity loss constrains the error between the predicted conductivity matrix and the actual conductivity matrix; the stage evolution consistency loss constrains the trend of predicted changes between adjacent construction stages to maintain consistency with the actual trend; the placement-conduction coupling consistency loss constrains the predicted conductivity result to maintain consistency with the empirical conductivity result derived from the predicted proppant placement result; and the physical constraint loss ensures that the predicted result conforms to fluid flow and particle transport laws.

3. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 2, characterized in that, The formula for calculating the total loss function is as follows: ； Among them, L S For proppant placement loss; L K This results in a loss of diversion capacity. L T This represents the consistency loss during stage evolution. L C To mitigate the loss of consistency in the flow-guiding coupling; L phy Loss due to physical constraints; , , , , These are the weighting coefficients for proppant placement loss, flow capacity loss, stage evolution consistency loss, placement-flow coupling consistency loss, and physical constraint loss, respectively.

4. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 3, characterized in that, The function of the proppant placement loss is: ； Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; S t The actual proppant placement matrix at time step t.

5. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 3, characterized in that, The function of the loss of the flow guiding capacity is: ； Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; K t Let be the actual flow control capability matrix at time step t.

6. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 3, characterized in that, The function of the stage evolution consistency loss is: ； Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; The proppant placement matrix predicted at time step t; The proppant placement matrix predicted at time step t-1; S t The actual proppant placement matrix at time step t; S t-1 The actual proppant placement matrix at time step t-1; Let be the flow capacity matrix predicted at time step t; This is the predicted flow capacity matrix at time step t-1; K t The actual flow capability matrix at time step t. K t-1 This is the actual flow capacity matrix at time step t-1.

7. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 3, characterized in that, The function of the consistency loss of the laid-out flow-guiding coupling is: ； Where T represents the total number of time steps corresponding to the construction phase; H and W are the height and width of the matrix, respectively; Let be the flow capacity matrix predicted at time step t; ； in, For the t-th time step The flow guidance capability matrix of each grid cell; Indicates the t-th time step. Predicted proppant placement value for each grid cell; Indicates the first Effective crack width of each grid cell; Indicates the first Normal closing stress of each grid cell; a, b, c, and d are coefficients.

8. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branch joints based on deep learning as described in claim 3, characterized in that, The function of the physical constraint loss is: ； in, These are the residual terms of the fluid control equations; This refers to the residual term of the particle motion equation; , These are the weighting coefficients; ； ； For fluid velocity; p represents pressure; For fluid density; For fluid viscosity; The velocity of the nth particle; , , , These represent drag force, buoyancy, gravity, and contact force, respectively. f is the force per unit volume of fluid; m is the mass of a single proppant particle; N is the number of sampling points; t is a continuous time variable.

9. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 3, characterized in that, The numerical model for proppant transport and placement covering multi-level branch cracks includes a main crack and first-, second-, and third-level branch cracks. The length, angle, and connection position of each branch crack to the main crack or the upper-level branch crack are all adjustable parameters. For branch cracks that do not exist in the simulation scheme, the corresponding positions in the model are left blank or given zero-concentration boundary conditions.

10. The prediction method for the combined proppant placement and flow conduction capacity of multi-level branched joints based on deep learning as described in claim 1, characterized in that, The recursive method yields the proppant placement matrix and conductivity matrix for each time step, including: Using the zero matrix or the initial stage state matrix as initial condition input, the proppant placement matrix and flow capacity matrix of the first time step are obtained; The prediction results of the previous time step are used as the conditional input for the next time step, and the proppant placement matrix and flow capacity matrix for each subsequent time step are obtained by recursion step by step.

Images