Crack morphology prediction method and device based on transformer, electronic equipment and storage medium

Abstract

The present application relates to the technical field of hydraulic fracturing, and is a fracture morphology prediction method and device based on a Transformer, an electronic device and a storage medium, which comprises obtaining key parameters of a well to be fractured in a set fracturing construction time period; inputting the key parameters of the well to be fractured in the set fracturing construction time period into a fracture morphology prediction model to obtain a fracture morphology prediction result, wherein an initial network model used for constructing the fracture morphology prediction model is set based on a Transformer network. The fracture morphology prediction model constructed by the present application performs parallel processing on feature extraction of static key parameters and dynamic key parameters, performs feature fusion after parallel processing, and enables static features and dynamic features to interact in a unified feature space based on the characteristics of the Transformer network, thereby improving the expression ability of the model for geological engineering coupling relationships and improving the prediction accuracy of the fracture morphology.

Description

Technical Field

[0001] This invention relates to the field of hydraulic fracturing technology, specifically a method, apparatus, electronic device, and storage medium for predicting fracture morphology based on Transformer. Background Technology

[0002] Hydraulic fracturing technology, as a key method for enhancing the production of unconventional underground energy resources, plays a crucial role in significantly improving reservoir permeability and productivity by inducing artificial fractures within the reservoir. However, the evolution of fracture geometry during fracturing directly affects not only the volume of the pressurized reservoir but also subsequent production capacity. Numerous studies have shown that fracture propagation behavior is subject to complex coupling effects from multiple factors, including reservoir heterogeneity, in-situ stress field distribution, fracturing fluid rheological properties, proppant migration patterns, and well completion design parameters. These factors collectively constitute a highly nonlinear, high-dimensional, spatiotemporally coupled physical system, posing significant challenges to the accurate characterization and quantitative analysis of fracture features.

[0003] Existing crack morphology prediction methods include: (1) Predicting fracture morphology based on idealized physical assumptions relies on idealized geomechanical assumptions and is difficult to accurately reflect the complexity of real reservoirs.

[0004] (2) Numerical simulation based on empirical formulas to predict fracture morphology. For example, existing patent document 1, publication number CN113821962B, discloses a method for predicting fracture morphology considering interlayer weak surfaces, including the following steps: establishing a three-dimensional geological model based on the obtained formation parameters of the actual interactive strata; establishing a rock seepage-stress coupling model based on the stress balance equation, continuity equation and boundary conditions; determining the fracture extension and propagation criteria; inputting the parameters of the target well area fracturing well to establish a three-dimensional fully coupled fracturing model; simulating the influence of different geological and engineering factors on fracture morphology. This invention establishes a full three-dimensional seepage-stress coupling model based on the finite element method, uses Cohesive pore pressure element to preset two types of fracture surfaces, considers interlayer weak surfaces or bedding weak surfaces, studies the influence of different geological and engineering factors on fracture morphology, and can quickly predict the fracture morphology of reservoirs containing interlayer weak surfaces during fracturing and guide the design of fracturing construction parameters. The applicability of this method is becoming increasingly limited. Such methods usually rely on idealized geomechanical assumptions, making it difficult to effectively integrate the high-frequency temporal dynamic information (such as pressure, displacement, sand ratio, etc.) contained in the fracturing construction curve and its complex coupling relationship with geomechanical parameters (such as geostress, rock brittleness, and the degree of development of natural fractures). Furthermore, they lack scalability and practicality in large-scale engineering scenarios such as multi-stage and multi-cluster fracturing.

[0005] (3) Utilize machine learning techniques to predict crack morphology, for example: Existing patent document 2, publication number CN121578405B, discloses a method, apparatus, and computing device for predicting the morphology of a hydraulic fracture network. The method includes: acquiring a source work area prediction model pre-trained based on work area data from a source work area; the source work area prediction model includes a source causal encoder and a morphology predictor; determining whether a target work area belongs to the source work area; if so, inputting the work area data of the target work area into the source work area prediction model to obtain the fracture morphology data output by the source work area prediction model; if not, using an adversarial fine-tuning algorithm to fine-tune the source causal encoder to obtain a target causal encoder; generating a target work area prediction model based on the target causal encoder and the morphology predictor; inputting the work area data of the target work area into the target work area prediction model to obtain the fracture morphology data output by the target work area prediction model. Existing patent document three, publication number CN121743699A, discloses a fracture morphology prediction method. This method constructs a database based on geological parameters, engineering parameters, and fracture morphology parameters of at least one oil well within a first preset time period. Each oil well includes at least one fracture formed using volumetric fracturing technology. The method involves obtaining logging curve parameters corresponding to at least one oil well. Based on these logging curve parameters, the geological parameters of any oil well within the first preset time period are expanded to update the database. A preset model is trained using the updated database to obtain a target model. The geological and engineering parameters of the oil well to be predicted within a second preset time period are then obtained. These parameters are input into the target model, and the output of the target model is determined as the fracture morphology parameters of the oil well to be predicted within the second preset time period, thereby improving the accuracy of fracture morphology prediction for the oil well. However, since the hydraulic fracturing process involves a large amount of multi-source heterogeneous data, including geological parameters, engineering parameters, microseismic monitoring results, downhole pressure / temperature records, and production dynamic data, existing technologies do not set up initial networks for the characteristics of different data types (static data and dynamic data). They cannot effectively extract and fuse the features of different data types, which makes it impossible for parameters such as reservoir properties, rock mechanics conditions, and construction parameters throughout the process to interact in a unified feature space. This hinders the model's ability to express the coupling relationship between geology and engineering and affects the prediction results of fracture morphology. Summary of the Invention

[0006] This invention provides a Transformer-based method, device, electronic device, and storage medium for predicting fracture morphology, overcoming the shortcomings of the prior art. It can effectively solve the problem that existing fracture morphology prediction methods cannot interact with parameters such as reservoir properties, rock mechanics conditions, and construction parameters throughout the entire process in a unified feature space, which hinders the model's ability to express the coupling relationship between geology and engineering and affects the fracture morphology prediction results.

[0007] One of the technical solutions of this invention is achieved through the following measures: a crack morphology prediction method based on Transformer, comprising: Obtain key parameters of the well to be fractured during the set fracture operation time period, including static key parameters and dynamic key parameters; The key parameters of the well to be fractured during the set fracturing operation period are input into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained by multiple learning samples on an initial network model. The initial network model is based on a Transformer network, extracts the static features of static key parameters and the dynamic features of dynamic key parameters and fuses them. From the fused features, the fracture response features corresponding to different time steps are extracted, and the fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

[0008] The following are further optimizations and / or improvements to the above-mentioned technical solution: The initial network model described above includes an encoding unit, a decoding unit, and a prediction unit; The encoding unit includes a static feature extraction module, a dynamic feature extraction module, and a fusion encoding module. The static feature extraction module and the dynamic feature extraction module extract static features and dynamic features from static key parameters and dynamic key parameters respectively in parallel. The fusion encoding module includes a fusion module and a multi-layer Transformer encoder. The fusion module places the static features at the beginning of the input sequence and then concatenates the dynamic features of each time step thereafter to obtain a fused input sequence. The multi-layer Transformer encoder interacts the static features and dynamic features in the fused input sequence in the same feature space to obtain fused features. The decoding unit extracts the crack response features corresponding to each time step based on learnable query features and fused features, and outputs the query features corresponding to each time step. The prediction unit uses a multilayer perceptron to perform regression calculations on the query features output by the decoding unit, and outputs a time series of predicted fracture geometric parameters as a function of the fracturing process. The time series includes the predicted fracture geometric parameters for each time step.

[0009] The construction process of the above crack morphology prediction model includes: Multiple learning samples were obtained and divided into training sample set, validation sample set and test sample set according to the proportion. Each learning sample included the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step during a certain fracturing operation period. The fracture geometry parameters included fracture length, fracture width and fracture height, and the key parameters included static key parameters and dynamic key parameters. The initial network model is trained using the training sample set. A loss function is introduced during training, and training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. The trained crack morphology prediction model is validated and its hyperparameters are adjusted using a validation sample set. The optimized crack morphology prediction model is then tested and evaluated using a test sample set, and a crack morphology prediction model that meets the test evaluation requirements is output.

[0010] The loss function is shown below: in, B Batch size; T For time steps; For the first b The learning sample at the th t Predicted values ​​of crack geometric parameters at each time step; For the b-th learning sample at the th t The true values ​​of the crack geometric parameters at each time step.

[0011] The process of obtaining the above learning samples includes: Multiple historical samples were obtained, each of which included the identifiers and candidate parameters of fracture geometry parameters at each time step during a specific fracturing operation period. The candidate parameters included static and dynamic candidate parameters. Static candidate parameters were parameters that did not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that were considered constant. Dynamic candidate parameters were parameters that changed with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Pearson correlation analysis was used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; Based on the parameter types of static and dynamic key parameters, each historical sample is reprocessed to obtain the corresponding learning sample.

[0012] The second technical solution of the present invention is achieved through the following measures: a crack morphology prediction device based on Transformer, comprising: The parameter acquisition unit acquires key parameters of the well to be fractured during the set fracture construction time period, including static key parameters and dynamic key parameters. The prediction unit inputs key parameters of the well to be fractured during the set fracturing operation period into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained on an initial network model using multiple learning samples. The initial network model is based on a Transformer network, extracts static features of static key parameters and dynamic features of dynamic key parameters and fuses them. From the fused features, fracture response features corresponding to different time steps are extracted, and fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

[0013] The following are further optimizations and / or improvements to the above-mentioned technical solution: The above also includes a model building unit for constructing a crack morphology prediction model, including: Multiple learning samples were obtained and divided into training sample set, validation sample set and test sample set according to the proportion. Each learning sample included the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step during a certain fracturing operation period. The fracture geometry parameters included fracture length, fracture width and fracture height, and the key parameters included static key parameters and dynamic key parameters. The initial network model is trained using the training sample set. A loss function is introduced during training, and training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. The trained crack morphology prediction model is validated and its hyperparameters are adjusted using a validation sample set. The optimized crack morphology prediction model is then tested and evaluated using a test sample set, and a crack morphology prediction model that meets the test evaluation requirements is output.

[0014] The above also includes a sample acquisition unit for acquiring training samples, including: Multiple historical samples were obtained, each of which included the identifiers and candidate parameters of fracture geometry parameters at each time step during a specific fracturing operation period. The candidate parameters included static and dynamic candidate parameters. Static candidate parameters were parameters that did not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that were considered constant. Dynamic candidate parameters were parameters that changed with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Pearson correlation analysis was used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; Based on the parameter types of static and dynamic key parameters, each historical sample is reprocessed to obtain the corresponding learning sample.

[0015] The third technical solution of the present invention is achieved through the following measures: an electronic device, characterized in that it includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement steps such as those in the Transformer-based crack morphology prediction method.

[0016] The fourth technical solution of the present invention is achieved by the following measures: a storage medium, characterized in that the storage medium stores a computer program that can be read by a computer, the computer program being configured to execute steps such as those in the crack morphology prediction method based on Transformer when running.

[0017] This invention introduces a fracture morphology prediction model based on a Transformer network, which improves the accuracy of fracture morphology prediction. Specifically, the fracture morphology prediction model adopts a structure of encoding unit, decoding unit, and prediction unit. The encoding unit performs parallel feature extraction on static and dynamic key parameters, and then performs feature fusion after parallel processing. Then, a self-attention mechanism is used to perform global correlation modeling on the fused static and dynamic features, enabling the model to automatically learn the dependencies between different input variables, especially to capture the pumping response relationship between long-distance time steps, overcoming the information decay problem that traditional sequence models are prone to in long sequence modeling. The decoding unit extracts fracture response information at each time step from the fused features through learnable query features, thereby realizing time-resolved fracture morphology prediction. This is beneficial for describing the dynamic evolution of fractures throughout the fracturing process, improving the model's adaptability to different construction systems and reservoir conditions. The multi-head attention structure extracts fracture propagation-related information in parallel from multiple feature subspaces, which can simultaneously characterize the influence of different geological parameters, engineering parameters, and time stages on fracture length, fracture width, and fracture height, improving the model's ability to learn complex nonlinear relationships and multi-factor coupling relationships. Attached Figure Description

[0018] Appendix Figure 1 This is a schematic diagram of an implementation environment provided for an embodiment of the present invention.

[0019] Appendix Figure 2 This is a schematic diagram of the crack morphology prediction method based on Transformer provided in an embodiment of the present invention.

[0020] Appendix Figure 3 This is a schematic diagram of the crack morphology prediction model construction method provided in an embodiment of the present invention.

[0021] Appendix Figure 4 This is a schematic diagram of the initial network model structure provided in an embodiment of the present invention.

[0022] Appendix Figure 5 This is a schematic diagram of the learning sample acquisition method provided in an embodiment of the present invention.

[0023] Appendix Figure 6 The graphs showing the trajectory changes of training loss and verification loss for crack prediction with training rounds are provided in the embodiments of the present invention. (a) is the curve showing the change of training loss and verification loss for crack length prediction with training rounds, (b) is the curve showing the change of training loss and verification loss for crack width prediction with training rounds, and (c) is the curve showing the change of training loss and verification loss for crack height prediction with training rounds.

[0024] Appendix Figure 7 This is a pumping timing diagram provided for an embodiment of the present invention.

[0025] Appendix Figure 8 The prediction and evaluation result curves provided in the embodiments of the present invention are shown in (a) for the prediction and evaluation result curve of crack length, (b) for the prediction and evaluation result curve of crack width, and (c) for the prediction and evaluation result curve of crack height.

[0026] Appendix Figure 9 This is a schematic diagram of a crack morphology prediction device based on Transformer provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0028] Those skilled in the art will understand that, unless specifically stated otherwise, in the embodiments of the present invention, a "module" or "unit" refers to a computer program or part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0029] In addition, in the embodiments of the present invention, "multiple" refers to two or more, and "first" and "second" are used to distinguish descriptions and should not be construed as implying relative importance.

[0030] This invention provides a Transformer-based fracture morphology prediction method. The method acquires key parameters of the well to be fractured during a set fracturing operation period, including static and dynamic key parameters. These key parameters are input into a fracture morphology prediction model to obtain the predicted fracture morphology results. The model is trained using multiple training samples on an initial network model based on a Transformer network. Static features of the static key parameters and dynamic features of the dynamic key parameters are extracted and fused. Fracture response features corresponding to different time steps are extracted from the fused features, and fracture geometric parameters corresponding to each time step are obtained based on these fracture response features.

[0031] The method provided in this embodiment of the invention may involve artificial intelligence (AI) technology and may be implemented based on artificial intelligence technology, such as using deep learning to train a corresponding model using samples.

[0032] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0033] Deep learning (DL) specifically refers to machine learning based on deep neural network models and methods. It has developed from statistical machine learning, artificial neural network algorithms, and other algorithms, combined with the advancements in big data and computing power. The most important technical feature of deep learning is its ability to automatically extract features.

[0034] The aforementioned machine learning and deep learning typically include techniques such as neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0035] In deep learning, the loss function is used to predict the target value by comparing the predicted value with the target value. This is done by updating the weight vector of each layer of the neural network based on the difference between the two values ​​(usually with an initialization process before the first update, where parameters are pre-configured for each layer) until the network can predict the target value or a value very close to it. Therefore, deep learning requires pre-defining "how to compare the difference between the predicted value and the target value," which is the loss function.

[0036] As attached Figure 1 The diagram illustrates an implementation environment provided by an embodiment of the present invention. This implementation environment may include: training equipment and usage equipment.

[0037] Both the training equipment and the equipment used are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, tablet computer, PC (Personal Computer) or other electronic devices; or, the computer device is a server, which can be a single server, a server cluster composed of multiple servers, or a cloud computing service center. This embodiment of the invention does not limit this.

[0038] Training equipment refers to computer equipment capable of network training and learning. Optionally, the training equipment has the ability to acquire an initial network model and train and learn it according to application requirements. For example, the training equipment acquires an initial network model from other devices via the network and then trains it using training samples according to application requirements so that the initial network model can obtain fracture morphology prediction values. Optionally, the training equipment has the ability to build an initial network model. It can build an initial network model itself according to application requirements and then train and learn it. For example, in order to obtain fracture morphology prediction values ​​based on key parameters of the well to be fractured during a set fracturing operation period, the training equipment builds an initial network model itself and then trains and learns it using samples according to application requirements.

[0039] The device used refers to a computer device that has the requirement to use an initial network model. Optionally, the device used can obtain the initial network model from other devices through the network according to the application requirements. For example, if the device used has the requirement to predict crack morphology, it can obtain the initial network model for training and learning to predict crack morphology from other devices through the network, and use the initial network model to predict crack morphology.

[0040] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0041] Example 1: As shown in the attached document Figure 2 As shown in the figure, this invention discloses a crack morphology prediction method based on Transformer, including: Step S110: Obtain the key parameters of the well to be fractured during the set fracture construction time period, including static key parameters and dynamic key parameters. Step S120: Input the key parameters of the well to be fractured during the set fracturing operation time period into the fracture morphology prediction model to obtain the fracture morphology prediction result. The fracture morphology prediction model is trained by using multiple learning samples to train the initial network model. The initial network model is based on the Transformer network setting, extracts the static features of the static key parameters and the dynamic features of the dynamic key parameters and fuses them. From the fused features, the fracture response features corresponding to different time steps are extracted, and the fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

[0042] In this embodiment, the key parameters include static key parameters and dynamic key parameters. The parameter types of both static and dynamic key parameters are obtained by screening from static and dynamic candidate parameters using correlation analysis. Static candidate parameters are parameters that do not change with the fracturing operation time, and may include geological parameters, rock mechanics parameters, and engineering parameters that are considered constant. Dynamic candidate parameters are parameters that do not change with the fracturing operation time, and may include pumping timing parameters that change with the fracturing operation process.

[0043] Specifically, the static key parameter is Xs, and the dynamic key parameter is a time series of uniform length. , where T is the standardized time step.

[0044] This invention discloses a crack morphology prediction method based on Transformer. It introduces a crack morphology prediction model based on Transformer network, performs parallel feature extraction on static and dynamic key parameters, performs feature fusion after parallel processing, and then, based on the characteristics of Transformer network, enables static and dynamic features to interact in a unified feature space, thereby improving the model's ability to express the coupling relationship between geology and engineering, and thus improving the accuracy of crack morphology prediction.

[0045] Example 2: As shown in the attached document Figure 3 As shown, the embodiments of the present invention are further optimizations of the above embodiments, wherein the process of constructing the crack morphology prediction model includes: Step S210: Obtain multiple learning samples and divide them into training sample set, validation sample set and test sample set according to the proportion. Each learning sample includes the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step in a certain fracturing construction period. The fracture geometry parameters include fracture length, fracture width and fracture height, and the key parameters include static key parameters and dynamic key parameters. In this embodiment, multiple learning samples are obtained. Each learning sample is an identifier and key parameter of the fracture geometry parameters of a historical fracturing well at each time step during a certain fracturing operation period. The specific key parameters include static key parameters during the fracturing operation period and dynamic key parameters at each time step during the fracturing operation period.

[0046] Step S220: Train the initial network model using the training sample set. During training, a loss function is introduced. Training ends when the training stopping condition is met, and a crack morphology prediction model is obtained. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. In this step, the loss function is set as needed, and can be, but is not limited to, mean squared error (MSE) to measure the difference between the predicted and actual values ​​of the crack geometry parameters: in, B Batch size; T For time steps; For the first b The learning sample at the th t Predicted values ​​of crack geometric parameters at each time step; For the b-th learning sample at the th tThe true values ​​of the crack geometric parameters at each time step.

[0047] This step updates the network parameters during training using the backpropagation algorithm until the training stopping condition is met, at which point a trained crack morphology prediction model is obtained. The training stopping condition includes the loss function value stabilizing or meeting the maximum number of iterations.

[0048] Step S230: Validate the trained crack morphology prediction model using the validation sample set and adjust the model hyperparameters. Test the optimized crack morphology prediction model using the test sample set and output a crack morphology prediction model that meets the test evaluation requirements.

[0049] In this embodiment, the test evaluation metrics may include, but are not limited to: Root Mean Square Error (RMSE): Mean Absolute Percentage Error (MAPE): Coefficient of determination (R) 2 ): in, These are the true values ​​of the crack's geometric parameters; These are the predicted values ​​for the crack's geometric parameters; This is the mean of the true values ​​of all crack geometric parameters.

[0050] Example 3: As shown in the attached document Figure 4 As shown, the embodiments of the present invention are further optimizations of the above embodiments, wherein the initial network model includes an encoding unit, a decoding unit, and a prediction unit, specifically: The encoding unit is used for feature extraction and fusion of static and dynamic key parameters. It includes a static feature extraction module, a dynamic feature extraction module, and a fusion encoding module. Specifically: Both the static feature extraction module and the dynamic feature extraction module are fully connected layers. Static key parameters and dynamic key parameters are mapped using the fully connected layers to obtain static features and dynamic features, respectively. The fusion coding module includes a fusion module and a multi-layer Transformer encoder, specifically: The fusion module adopts a static feature fusion labeling strategy, which places static features at the beginning of the input sequence and then concatenates dynamic features from each time step thereafter to obtain the fused input sequence. in, It is a static feature; For the first tTo preserve the temporal relationship of the dynamic features, positional encoding is added to the dynamic features at each time step. The multi-layer Transformer encoder includes a multi-head self-attention mechanism layer and a feedforward neural network layer. Each layer is equipped with residual connections and layer normalization. The multi-layer Transformer encoder interacts with the static and dynamic features in the input sequence in the same feature space to obtain fused features.

[0051] In the self-attention calculation of a multi-layer Transformer encoder, the query vector Q, key vector K, and value vector V all come from the fused input sequence, and therefore can be represented as: The attention calculation process is as follows: in, d k The dimension of the key vector.

[0052] By using the multi-head self-attention mechanism of the multi-layer Transformer encoder, static features and dynamic features (i.e., static geological parameters, rock mechanics parameters, and dynamic pumping parameters) can interact in the same feature space, thereby obtaining fused features M.

[0053] The decoding unit extracts the crack response features corresponding to each time step based on learnable query features and fused features, and outputs the query features corresponding to each time step. The decoding unit is a Transformer decoder, which includes a self-attention layer, a cross-attention layer and a feedforward neural network. Residual connections and layer normalization are set in each layer.

[0054] It should be noted that in the self-attention layer, the query vector, key vector, and value vector all come from the query features of the current layer, and are used to establish the dependencies between query features at different time steps, i.e.: Q=K=V=Q L In the cross-attention layer, the query vector comes from the query features on the decoder side, and the key and value vectors come from the fused features output by the encoder. These are used to enable the query features at each time step to retrieve features related to crack morphology prediction from the fused features. in, represents the query features updated by self-attention; M represents the fused features output by the encoding unit.

[0055] The prediction unit uses a multilayer perceptron to perform regression calculations on the query features output by the decoding unit, outputting a time series of predicted fracture geometric parameters as a function of the fracturing process. The time series includes the predicted fracture geometric parameters for each time step, where the predicted fracture geometric parameters for the t-th time step are shown below: in, The first t The crack length, crack width, and crack height are predicted at each time step.

[0056] In this embodiment, the fracture morphology prediction model adopts a structure of encoding unit, decoding unit, and prediction unit. The encoding unit performs parallel feature extraction on static and dynamic key parameters, performs feature fusion after parallel processing, and then uses a self-attention mechanism to perform global correlation modeling on the fused static and dynamic features. This enables the model to automatically learn the dependencies between different input variables, especially to capture the pumping response relationship between long-distance time steps, overcoming the information decay problem that traditional sequence models are prone to in long sequence modeling. The decoding unit extracts fracture response information of each time step from the fused features through learnable query features, thereby realizing time-resolved fracture morphology prediction. This is beneficial for describing the dynamic evolution of fractures throughout the fracturing process and improves the model's adaptability to different construction regimes and reservoir conditions. The multi-head attention structure extracts fracture propagation-related information in parallel from multiple feature subspaces, which can simultaneously characterize the influence of different geological parameters, engineering parameters, and time stages on fracture length, fracture width, and fracture height, improving the model's ability to learn complex nonlinear relationships and multi-factor coupling relationships.

[0057] Example 4: As shown in the appendix Figure 5 As shown, the embodiments of the present invention are further optimizations of the above embodiments, wherein the process of obtaining learning samples includes: Step S310: Obtain multiple historical samples, each of which includes the identifier and candidate parameters of the fracture geometry parameters at each time step during a certain fracturing operation period of the historical fracturing well. The candidate parameters include static candidate parameters and dynamic candidate parameters. Static candidate parameters are parameters that do not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that are considered constant. Dynamic candidate parameters are parameters that change with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Step S320: Pearson correlation analysis is used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; This screening process uses crack length, crack width, and crack height as independent output variables to perform Pearson correlation analysis to complete the screening. Specifically: For the j Candidate input variables X j Calculate its relationship with the crack length respectively. Y L Crack width Y W and crack height Y H The Pearson correlation coefficient between them was obtained. ; in, Y k Output variables for crack geometry. k Desirable L , W , H These correspond to crack length, crack width, and crack height, respectively. X ij For the first i In the nth sample j The values ​​of the candidate input variables; Y k，i For the first i The corresponding sample of the th sample k Crack-like geometric parameters; where, The mean of the corresponding variable across all samples; Each candidate input variable corresponds to three correlation coefficients. To comprehensively evaluate the overall influence of the candidate input variable on the crack geometric parameters, the average absolute correlation coefficient is used as the comprehensive correlation index. Meanwhile, to avoid omitting variables that have a significant impact on a single crack geometric parameter, the maximum absolute correlation coefficient is introduced as an auxiliary indicator: Based on the above process, correlation heatmaps were established between candidate input variables and crack length, crack width, and crack height, respectively, and according to... The candidate input variables are comprehensively ranked and screened based on their engineering and physical significance. Variables with high overall correlation, or those with moderate overall correlation but significant impact on a certain crack geometric parameter and clear physical significance, are retained; variables with weak correlation, unclear physical significance, or high redundancy with other variables are eliminated.

[0058] Step S330: Based on the parameter types of static and dynamic key parameters, reprocess each historical sample to obtain the corresponding learning sample.

[0059] Example 5: This example verifies the effectiveness of the crack morphology prediction model of the present invention. Specifically: The dataset containing 20,000 training samples was used for testing. It was randomly divided into training, validation and test sets in a ratio of 7:2:1 (i.e., 14,000 / 4,000 / 2,000). The training set was used for model parameter optimization, the validation set was used for hyperparameter tuning and early stopping control, and the test set was used for unbiased performance evaluation.

[0060] The initial network model is trained using the training sample set, with a loss function introduced during training. Training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The trained crack morphology prediction model is then validated and tested using the validation and test sample sets, respectively. The model parameters of the crack morphology prediction model are optimized, and a crack morphology prediction model that meets the test evaluation requirements is output.

[0061] Appendix Figure 6 The training and validation losses for predicting crack length, crack width, and crack height are shown as curves over training epochs. An epoch refers to the process of the model learning the entire training dataset. Early stopping is a method used in model training to prevent overfitting; training is terminated early when performance on the validation dataset no longer improves, rather than mechanically running all preset epochs. All target variables show a rapid decrease in loss in the initial few epochs, followed by a stable convergence phase. The training and validation losses are highly consistent, showing no significant divergence, indicating effective model regularization and no significant overfitting. The model reaches near-optimal performance within approximately 30 to 60 training epochs, demonstrating good training stability and computational efficiency.

[0062] After convergence, the model's time-domain estimation capability throughout the hydraulic fracturing process was evaluated, using the same pumping scheme, as shown in the attached figure. Figure 7 As shown, the mud flow rate, proppant concentration, and fluid volume are also labeled.

[0063] As attached Figure 8 As shown, quantitative analysis reveals that the coefficients of determination (R²) for all three indicators are... 2The values ​​of all values ​​exceeded 0.95, and the MAPE values ​​were all below 10%, indicating that the Transformer model can reliably reconstruct the dynamic changes in fracture geometry over time during hydraulic fracturing. These results confirm that the proposed framework can consistently correlate dynamic pumping conditions with fracture propagation behavior in a physically consistent manner.

[0064] Example 6: As shown in the appendix Figure 9 As shown, this embodiment of the invention discloses a crack morphology prediction device based on Transformer, comprising: The parameter acquisition unit acquires key parameters of the well to be fractured during the set fracture construction time period, including static key parameters and dynamic key parameters. The prediction unit inputs key parameters of the well to be fractured during the set fracturing operation period into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained on an initial network model using multiple learning samples. The initial network model is based on a Transformer network, extracts static features of static key parameters and dynamic features of dynamic key parameters and fuses them. From the fused features, fracture response features corresponding to different time steps are extracted, and fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

[0065] Example 7: This embodiment of the invention discloses a crack morphology prediction device based on Transformer, comprising: The sample acquisition unit, used to acquire training samples, includes: Multiple historical samples were obtained, each of which included the identifiers and candidate parameters of fracture geometry parameters at each time step during a specific fracturing operation period. The candidate parameters included static and dynamic candidate parameters. Static candidate parameters were parameters that did not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that were considered constant. Dynamic candidate parameters were parameters that changed with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Pearson correlation analysis was used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; Based on the parameter types of static and dynamic key parameters, each historical sample is reprocessed to obtain the corresponding learning sample.

[0066] Model building units, used to construct crack morphology prediction models, include: Multiple learning samples were obtained and divided into training sample set, validation sample set and test sample set according to the proportion. Each learning sample included the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step during a certain fracturing operation period. The fracture geometry parameters included fracture length, fracture width and fracture height, and the key parameters included static key parameters and dynamic key parameters. The initial network model is trained using the training sample set. A loss function is introduced during training, and training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. The trained crack morphology prediction model is validated and its hyperparameters are adjusted using a validation sample set. The optimized crack morphology prediction model is then tested and evaluated using a test sample set, and a crack morphology prediction model that meets the test evaluation requirements is output.

[0067] The parameter acquisition unit acquires key parameters of the well to be fractured during the set fracture operation time period, including static key parameters and dynamic key parameters.

[0068] The prediction unit inputs key parameters of the well to be fractured during the set fracturing operation period into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained on an initial network model using multiple learning samples. The initial network model is based on a Transformer network, extracts static features of static key parameters and dynamic features of dynamic key parameters and fuses them. From the fused features, fracture response features corresponding to different time steps are extracted, and fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

[0069] Example 8: This embodiment of the invention discloses a storage medium storing a computer program that can be read by a computer. The computer program is configured to execute steps in a Transformer-based crack morphology prediction method at runtime.

[0070] The aforementioned storage media may include, but are not limited to, USB flash drives, read-only memory, portable hard drives, magnetic disks, optical disks, and other media capable of storing computer programs.

[0071] Example 9: This embodiment of the invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps in the Transformer-based crack morphology prediction method.

[0072] The processor described above can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, portable hard drives, magnetic disks, or optical disks.

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] The above content is only a specific embodiment of the present invention, which has strong adaptability and implementation effect. However, the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention. Therefore, equivalent changes made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A crack morphology prediction method based on Transformer, characterized in that, include: Obtain key parameters of the well to be fractured during the set fracture operation time period, including static key parameters and dynamic key parameters; The key parameters of the well to be fractured during the set fracturing operation period are input into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained on an initial network model using multiple training samples. The initial network model is based on a Transformer network, extracting static features of static key parameters and dynamic features of dynamic key parameters. The static features are placed at the beginning of the input sequence, and then the dynamic features of each time step are concatenated sequentially to obtain the fused input sequence. , It is a static feature. All are dynamic features. Crack response features corresponding to different time steps are extracted from the fused features, and crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to different time steps.

2. The crack morphology prediction method based on Transformer according to claim 1, characterized in that, The initial network model includes an encoding unit, a decoding unit, and a prediction unit; The coding unit includes a static feature extraction module, a dynamic feature extraction module, and a fusion coding module; the static feature extraction module and the dynamic feature extraction module extract static features and dynamic features respectively from static key parameters and dynamic key parameters in parallel; The fusion coding module includes a fusion module and a multi-layer Transformer encoder. The fusion module places the static features at the beginning of the input sequence and then concatenates the dynamic features of each time step thereafter to obtain the fused input sequence. The multi-layer Transformer encoder interacts the static and dynamic features in the fused input sequence in the same feature space to obtain the fused features. The decoding unit extracts the crack response features corresponding to each time step based on learnable query features and fused features, and outputs the query features corresponding to each time step. The prediction unit uses a multilayer perceptron to perform regression calculations on the query features output by the decoding unit, and outputs a time series of predicted fracture geometric parameters as a function of the fracturing process. The time series includes the predicted fracture geometric parameters for each time step.

3. The crack morphology prediction method based on Transformer according to claim 1 or 2, characterized in that, The construction process of the crack morphology prediction model includes: Multiple learning samples were obtained and divided into training sample set, validation sample set and test sample set according to the proportion. Each learning sample included the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step during a certain fracturing operation period. The fracture geometry parameters included fracture length, fracture width and fracture height, and the key parameters included static key parameters and dynamic key parameters. The initial network model is trained using the training sample set. A loss function is introduced during training, and training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. The trained crack morphology prediction model is validated and its hyperparameters are adjusted using a validation sample set. The optimized crack morphology prediction model is then tested and evaluated using a test sample set, and a crack morphology prediction model that meets the test evaluation requirements is output.

4. The crack morphology prediction method based on Transformer according to claim 3, characterized in that, The loss function is as follows: in, B Batch size; T For time steps; For the first b The learning sample at the th t Predicted values ​​of crack geometric parameters at each time step; For the b-th learning sample at the th t The true values ​​of the crack geometric parameters at each time step.

5. The crack morphology prediction method based on Transformer according to claim 1 or 2, characterized in that, The process of acquiring the learning samples includes: Multiple historical samples were obtained, each of which included the identifiers and candidate parameters of fracture geometry parameters at each time step during a specific fracturing operation period. The candidate parameters included static and dynamic candidate parameters. Static candidate parameters were parameters that did not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that were considered constant. Dynamic candidate parameters were parameters that changed with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Pearson correlation analysis was used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; Based on the parameter types of static and dynamic key parameters, each historical sample is reprocessed to obtain the corresponding learning sample.

6. A Transformer-based crack morphology prediction device applying the method described in any one of claims 1 to 5, characterized in that, include: The parameter acquisition unit acquires key parameters of the well to be fractured during the set fracture construction time period, including static key parameters and dynamic key parameters. The prediction unit inputs key parameters of the well to be fractured during the set fracturing operation period into the fracture morphology prediction model to obtain the fracture morphology prediction results. The fracture morphology prediction model is trained on an initial network model using multiple learning samples. The initial network model is based on a Transformer network, extracts static features of static key parameters and dynamic features of dynamic key parameters and fuses them. From the fused features, fracture response features corresponding to different time steps are extracted, and fracture geometric parameters corresponding to each time step are obtained based on the fracture response features corresponding to different time steps.

7. The crack morphology prediction device based on Transformer according to claim 6, characterized in that, It also includes model building units for constructing crack morphology prediction models, including: Multiple learning samples were obtained and divided into training sample set, validation sample set and test sample set according to the proportion. Each learning sample included the identifier and key parameters of the fracture geometry parameters of the historical fracturing well at each time step during a certain fracturing operation period. The fracture geometry parameters included fracture length, fracture width and fracture height, and the key parameters included static key parameters and dynamic key parameters. The initial network model is trained using the training sample set. A loss function is introduced during training, and training ends when the training stopping condition is met, resulting in a crack morphology prediction model. The initial network model is based on the Transformer network. Static features of static key parameters and dynamic features of dynamic key parameters are extracted and fused. Crack response features corresponding to each time step are extracted from the fused features. Crack geometric parameters corresponding to each time step are obtained based on the crack response features corresponding to each time step. The trained crack morphology prediction model is validated and its hyperparameters are adjusted using a validation sample set. The optimized crack morphology prediction model is then tested and evaluated using a test sample set, and a crack morphology prediction model that meets the test evaluation requirements is output.

8. The Transformer-based crack morphology prediction device according to claim 6 or 7, characterized in that, It also includes a sample acquisition unit for acquiring training samples, including: Multiple historical samples were obtained, each of which included the identifiers and candidate parameters of fracture geometry parameters at each time step during a specific fracturing operation period. The candidate parameters included static and dynamic candidate parameters. Static candidate parameters were parameters that did not change with the fracturing operation time, including geological parameters, rock mechanics parameters, and engineering parameters that were considered constant. Dynamic candidate parameters were parameters that changed with the fracturing operation time, including proppant displacement, fracturing fluid injection volume, fluid viscosity, and proppant characteristics. Pearson correlation analysis was used to screen static and dynamic candidate parameters to obtain the parameter types of static and dynamic key parameters; Based on the parameter types of static and dynamic key parameters, each historical sample is reprocessed to obtain the corresponding learning sample.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps of the method as claimed in any one of claims 1 to 5.

10. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, the computer program being configured to execute the steps of the method as described in any one of claims 1 to 5 when it is run.

Images