Intelligent prediction method for fracture pressure of compact conglomerate reservoir based on improved Transform architecture

Abstract

The invention discloses an intelligent prediction method for fracture pressure of a compact conglomerate reservoir based on an improved Transform architecture, and relates to the technical field of combination of deep learning and oil extraction. Comprising the following steps: acquiring fracture pressure data of fractured wells in the same block and logging parameters related to the fracture pressure, matching the fracture pressure with the logging parameters, and preprocessing the data to obtain initial data; based on a CAE-SOM self-organizing mapping neural network, clustering the initial data to obtain sample categories under different geological conditions; based on a generative adversarial network, sample categories under any geological condition are expanded to obtain a training set; and in combination with an attention mechanism, training the Transform model, and predicting the fracture pressure of the block reservoir by using the trained Transform prediction model. According to the method, the limited fracture pressure data can be accurately utilized for prediction, so that a more efficient method is provided for providing powerful support for fracture design and segmented clustering optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of combining deep learning with oil production, and specifically to an intelligent prediction method for fracture pressure of dense conglomerate reservoirs based on an improved Transformer architecture. Background Art

[0002] The tight conglomerate reservoir has strong heterogeneity and large differences in rock fracture pressure, which leads to uneven crack expansion during staged fracturing of horizontal wells, directly affecting the production effect of horizontal wells. With the help of FSI production profile data, the true value of the fracture pressure of each cluster of horizontal wells can be calibrated with the logging data. However, the high cost of FSI and the limited number of real samples for calibration bring challenges to the accurate prediction of fracture pressure.

[0003] Existing machine learning methods have strict data requirements. When faced with limited data sets, these methods often have difficulty in building models with strong generalization capabilities, especially when dealing with complex geological conditions with significant heterogeneity such as tight conglomerate reservoirs. Since the internal properties of such reservoirs vary greatly, limited data samples may not be able to fully reflect their diversity and complexity, resulting in poor performance of the trained models in practical applications, especially in critical operations such as fracturing design and segmentation and clustering optimization, which will be difficult to meet engineering needs due to lack of sufficient accuracy. Summary of the invention

[0004] To solve at least one of the above problems, the present invention proposes an intelligent prediction method for the fracture pressure of a tight conglomerate reservoir based on an improved Transformer architecture.

[0005] The technical solution of the present invention is: an intelligent prediction method for the fracture pressure of a dense conglomerate reservoir based on an improved Transformer architecture, comprising the following steps: S1. Obtain the fracture pressure data of the fractured wells in the same block and the logging parameters related to the fracture pressure, match the fracture pressure and the logging parameters, and pre-process the data to obtain the initial data; S2, based on CAE-SOM self-organizing map neural network, considering the heterogeneity of the reservoir, clustering the initial data to obtain sample categories under different geological conditions; S3. Based on the generative adversarial network, the sample categories under any geological conditions are expanded to obtain the training set; S4. Combining the global attention and local attention mechanisms, the Transformer model is trained using the training set under any geological condition, and finally the Transformer prediction model under any geological condition is obtained; S5. Determine the sample category of the target well, and use the Transformer prediction model corresponding to the sample category to predict its fracture pressure.

[0006] Beneficial effects: The method of the present invention clusters samples of the real values ​​of the fracture pressure, expands the sample set for the same type of samples through the generative adversarial network, and then establishes a prediction model for each type of expanded sample set, combining the advantages of the self-organizing map neural network and the generative adversarial network to effectively solve the above technical problems. By clustering and expanding the data of the acquired real values ​​of the fracture pressure, the present invention can accurately use limited fracture pressure data for prediction, thereby providing a more efficient method to provide strong support for fracturing design and segmented clustering optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 are the values ​​of depth, natural gamma, acoustic time difference and true resistivity of formation in different samples; Figure 2 are the values ​​of density, well diameter, compensated neutron and natural potential in different samples; Figure 3 is the value of rupture pressure in different samples; Figure 4 4 is a comparison chart of the model prediction results and the measured results of this embodiment. DETAILED DESCRIPTION

[0008] The specific implementation modes of the present invention will be described clearly and completely below with reference to examples and drawings. Obviously, the described examples are only some embodiments of the present invention, rather than all embodiments.

[0009] An intelligent prediction method for tight conglomerate reservoir fracture pressure based on an improved Transformer architecture comprises the following steps: S1. Obtain the fracture pressure data of the fractured wells in the same block and the logging parameters related to the fracture pressure, match the fracture pressure and the logging parameters, and pre-process the data to obtain the initial data; Specifically, in this embodiment, the logging parameters include depth, natural gamma, acoustic time difference, true resistivity of formation, density, well diameter, compensated neutron and natural potential. These parameters are relatively easy to obtain and can be obtained through logging while drilling.

[0010] The fracture pressure can be obtained by the following methods: recording the peak pressure during staged fracturing by observing the bottom hole pressure changes during the fracturing process; using FSI data to identify the fractured clusters, and matching the peak pressure with the fracture clusters according to the principle of principal stress magnitude to obtain the true value of the fracture pressure of the fracture cluster.

[0011] Finally, the logging parameters and fracture pressure of the fracturing cluster are matched.

[0012] For tight conglomerate reservoirs, multi-stage and multi-cluster fracturing is usually performed. The specific geological conditions of different fracturing stages and fracturing clusters are different, resulting in different fracture pressures. Therefore, we first need to obtain the fracture pressure of each fracturing cluster and the corresponding logging parameters, and match the fracture pressure and logging parameters.

[0013] However, in the actual production process, the obtained data may contain missing values ​​and outliers. In order to reduce the impact of these missing values ​​and outliers on the final prediction results, in this embodiment, the segmented regression interpolation method is used to correct them: the data corresponding to each parameter in the acquired logging while drilling parameters is used as a set of fitting data, and a univariate cubic regression curve is fitted. When the gap between the acquired value and the fitting value is too large, the acquired value is regarded as noise and replaced by the fitting value. In this embodiment, if the absolute difference between the acquired value and the fitting value is greater than 5%, it is considered that the gap is too large. Those skilled in the art can set different upper and lower limits according to actual conditions; for missing values ​​in the data, the average value of the group of data is used instead. Until the outlier processing of all data is completed; The processed data is normalized. Normalization is a conventional operation in the art. In this embodiment, the maximum-minimum normalization method is used to normalize the data, and the specific operation is not described in detail here.

[0014] At the same time, in this step, for the fracture pressure, the acquisition process of its real sample needs to use the FSI output profile logging data. However, those skilled in the art know that the cost of FSI output profile logging data is relatively high. Therefore, we only obtained a limited data set, such as Figure 1~Figure 3 As shown, a total of 50 real samples are selected in this embodiment. Figure 1 The values ​​of depth, natural gamma, acoustic time difference and true resistivity of formation in different samples are shown in turn. Figure 2 The values ​​of density, well diameter, compensated neutron and natural point in different samples are shown in turn. Figure 3 The values ​​of rupture pressure in different samples are shown.

[0015] S2. Based on the CAE-SOM self-organizing map neural network, the initial data is clustered considering the heterogeneity of the reservoir to obtain the sample clustering results under different geological conditions; For dense conglomerate reservoirs, the heterogeneity is strong. The stress difference of the rock is large in different fracturing sections or even different fracturing clusters, which leads to large differences in the fracture pressure. Therefore, if the conventional SOM self-organizing map neural network is directly used for clustering analysis, the clustering results will be distorted. For this reason, the inventor first uses the autoencoder to extract the effective heterogeneity characteristics of the dense conglomerate reservoir before clustering. Therefore, the specific operation of this step is as follows.

[0016] 1. Extracting effective heterogeneous features of tight conglomerate reservoirs using autoencoders: In this process, the logging parameters of each fracturing cluster are first combined into an input variable: [depth, density, resistivity, acoustic time difference, natural gamma, natural potential, compensated neutron, wellbore], where, is the number of the fracturing cluster, For different logging parameters.

[0017] 2. Then use the autoencoder to extract effective heterogeneous features. The autoencoder is divided into an encoder and a decoder.

[0018] 2.1 The encoder includes: Convolutional layer: used to extract local features from the input logging data. The logging data is a The matrix of is the total number of samples, is the number of LWD parameters.

[0019] Convolution kernel: choose a size of The convolution kernel, For different convolution kernels, the convolution kernel slides on the input data, performs weighted summation, and extracts local features in the data; Convolution operation: The convolution operation generates multiple feature maps, each feature map represents the response to a specific pattern in the input data.

[0020] The convolution operation is as follows:

[0021] in: For the convolution kernel Next The corresponding fracturing cluster samples Characteristic elements of LWD parameters, For input data, It is The weight of the convolution kernel, It is The bias term of the convolution kernel. is the relative position index within the convolution kernel, and are the offsets of the convolution kernel in the row and column directions, respectively. Indicates that in the row direction from Cluster Offset The row index after the position, Indicates that in the column direction from Logging parameter migration The row index after the position.

[0022] (3) Pooling layer: used to reduce the spatial dimension of the feature map and retain significant features. Maximum pooling is used as the pooling operation, which takes the maximum value of each local area.

[0023] The pooling operation is as follows:

[0024] in: For input data, is the pooling window size, Represents the output, which is the local maximum of the input.

[0025] Through pooling, the convolutional autoencoder gradually reduces the spatial dimension of the input data and extracts key information from the data.

[0026] Latent representation: Through the combination of multiple convolutional layers and pooling layers, the encoder compresses the input data from the original space to a low-dimensional latent space representation. This latent representation captures the key features and structural information of the data, which is crucial for describing the heterogeneity of tight conglomerate reservoirs. The dimension of the latent space is usually low, but it contains deep features extracted from the original data. The output data is .

[0027] 2.2 The decoder of the autoencoder includes: The decoder gradually restores the spatial resolution of the feature map through the deconvolution (transposed convolution) layer, and finally restores it to the same size as the input data. The deconvolution operation is similar to the convolution operation, but it maps the potential representation back to a high-dimensional space through learning. The mathematical formula of deconvolution is similar to the convolution operation, usually by upsampling the potential representation (such as the deconvolution operation) and then restoring it through the filter convolution.

[0028] Reconstruction error: The decoder finally outputs the reconstructed data , and with the input data By minimizing the difference between the input data and the reconstructed data (usually using the mean square error loss function ), the convolutional autoencoder can train effective feature representation:

[0029] In the formula, Represents the reconstructed data finally output by the decoder, Represents input data, represents the 2-norm.

[0030] The convolutional autoencoder is trained through the back-propagation algorithm, and the optimization goal is to minimize the error between the input data and the reconstructed data. Through training, the convolutional autoencoder can automatically learn effective features that represent reservoir heterogeneity.

[0031] Normalize the data retrieved through the convolutional autoencoder to its range;

[0032] In the formula, Indicates The value of a variable in a sample, Indicates the sample number.

[0033] In this embodiment, the output dimension of the autoencoder is set to 4, and those skilled in the art can set a suitable output dimension according to actual conditions.

[0034] 3. After extracting effective heterogeneous features through the autoencoder, the extracted effective heterogeneous features are brought into the SOM self-organizing map neural network for clustering.

[0035] The SOM self-organizing map neural network includes the following steps: (1) defining the size of the self-organizing map SOM competition layer and initializing the weight matrix, wherein each row in the weight matrix corresponds to a weight vector of a competition layer point in the competition layer; (2) Input sample: The feature vector extracted by the convolutional autoencoder As the input sample of SOM.

[0036] (3) Select a sample from the sample data set as the input vector, calculate the Euclidean distance between the sample and the weight matrix, find the nearest competitive layer point as the activation point, set its weight to 1, then the Euclidean distance between the sample vector and an element in the weight matrix is The calculation formula is:

[0037] In the formula, Indicates The input vector corresponding to the samples is represents the weight of the lth sample in the weight matrix, Indicates the sample number, Indicates the variable number, Represents the norm.

[0038] (4) The competitive layer point to which the vector with the closest distance belongs is the winning point. The other competitive layer points in the neighborhood are determined according to the coordinates and domain range of the winning point, and the weights of other competitive layer points in the neighborhood are updated. The weight update formula is as follows:

[0039] Represents the time step When the weight matrix The weight of the samples, Represents the time step When the weight matrix The weight of the samples.

[0040] The update formula of learning rate is:

[0041] In the formula, represents the updated learning rate, represents the initial learning rate, Indicates the maximum number of time steps.

[0042] (5) Repeat steps (3)-(4) until the number of iterations is met, input all sample data sets to obtain clustering results, and finally obtain Clustering results.

[0043] (6) Output geological categories: The trained SOM network generates a two-dimensional map in which each node represents a cluster with similar geological conditions. By observing the map, the distribution of different geological types can be intuitively seen, so that clusters with similar geological conditions can be classified into the same category. In this embodiment, based on the SOM network, the initial data is finally divided into three categories.

[0044] S3. Based on the generative adversarial network, the clustering results of samples under any geological conditions are expanded to obtain the training set; In step 2, we get The clustering results are shown in Figure 2. Since the acquisition process of the real samples of fracture pressure needs to rely on the FSI output profile logging data to observe, and the FSI cost is high, the number of real samples of fracture pressure is limited. Therefore, before predicting the fracture pressure, it is necessary to expand the samples in each cluster. GAN can grasp the complex distribution of data through unsupervised learning, thereby generating synthetic data with high quality and reliability. Therefore, the GAN algorithm is used as the generation algorithm of fracture pressure samples.

[0045] The specific steps include: (1) Design GAN structure: Generator: The task of the generator is to generate samples similar to the true data distribution based on the noise vector and class labels.

[0046] Input layer: noise vector : Sample from a standard normal distribution. This represents randomness and controls the diversity of the generated data; class labels : A uniquely encoded category label indicating which category of data the generator generates. Concatenated together as input to the generator.

[0047] Hidden layer: A series of fully connected layers or deconvolution layers that gradually map the input low-dimensional vector to the high-dimensional space of the data.

[0048] Output layer: Output a vector with the same characteristics as the original data.

[0049] Activation function: The hidden layer uses the ReLU activation function and the output layer uses the Sigmoid activation function to ensure that the values ​​of the generated data are within the appropriate range.

[0050] Discriminator: The task of the discriminator is to determine whether the input sample is real. Its goal is to output 1 for real data and 0 for generated data.

[0051] Input layer: input is a sample data , which can be a real sample or a fake sample generated by the generator; category label :It is used as a conditional input together with the input sample to help the discriminator know which category the data belongs to. : and category labels Concatenate and pass into the discriminator.

[0052] Hidden layers: Multiple fully connected layers or convolutional layers are used to extract features of the input data.

[0053] Output layer: Finally, it outputs a scalar, indicating whether the sample is real (1) or generated (0).

[0054] Activation function: The hidden layer uses the ReLU activation function, and the output layer uses the sigmoid activation function. The output is a probability value, indicating the probability that the input sample is real data.

[0055] (2) Training Generative Adversarial Network (GAN): The training process of GAN is achieved by alternating optimization of the generator and the discriminator.

[0056] Discriminator training: Training real data: Training real data: Use real data samples to train the discriminator so that it can correctly predict real samples as "real" (labeled as 1).

[0057]

[0058] here is the probability of the discriminator output, indicating the input sample The probability of belonging to the real data. To calculate the discriminator In real data The loss on for From the real data distribution The expected value of the sample.

[0059] Training generated samples: Use fake samples generated by the generator to train the discriminator so that it can correctly predict fake samples as "fake" (label 0).

[0060]

[0061] here is a fake sample generated by the generator, is the discriminator’s prediction for the generated sample. To calculate the discriminator In the fake sample The loss on express From the noise distribution The expected value of the sample.

[0062] Generator training: The goal of the generator is to confuse the discriminator into thinking that the generated fake samples are real. The loss function of the generator is optimized by maximizing the probability that the discriminator predicts the generated samples as "real".

[0063]

[0064] Here For the generator loss.

[0065] Training the generator: During training, fix the discriminator and update the generator so that the fake samples it generates can maximize the discriminator's prediction probability that it is a real sample.

[0066] The training process is carried out alternately: first train the discriminator and optimize it to distinguish between real samples and generated samples, then fix the discriminator and train the generator to make the samples it generates more realistic and can confuse the discriminator. Repeat these training steps until the samples generated by the generator can well simulate the distribution of real samples.

[0067] (3) Generate augmented data: Once the GAN training is completed, the trained generator can be used to generate new samples. Sampling in the middle, combined with category labels , generate new samples through the generator. That is, given noise and category labels as input, the generator will generate new data samples.

[0068] -

[0069] here is the input of the generator, Represents the new sample generated.

[0070] When the sample is expanded by using a generative adversarial network, the sample size is usually expanded by 2-10 times the initial sample size. Of course, those skilled in the art can select a suitable expansion endorsement according to the actual situation. In this embodiment, the sample size is expanded by 5 times.

[0071] S4, combining global attention and local attention mechanisms, using the training set under any geological condition to train the Transformer model, and finally obtaining the Transformer prediction model under any geological condition, and using the Transformer prediction model to predict the fracture pressure of the reservoir in this block. This step specifically includes the following sub-steps: (1) Randomly select a training set and transform its data into vectors in a high-dimensional space; let ,in is a arrive The embedding function of . Indicates the logging while drilling parameters exist Encoding in dimensional space, value is called the dimension of the model.

[0072] In this step, the logging parameters are encoded into a number of new variables, among which:

[0073] (2) Add position coding information to the rupture pressure data in the high-level adjacent space; the position coding method is shown in the following formula:

[0074] In the formula, represents the positional encoding of even dimensions, represents the positional encoding of odd dimensions, Represents the model dimension, Represents a position in the input sequence.

[0075] (3) Using Q, K, and V, we calculate local attention and global attention respectively and combine their calculation results into multi-head attention to obtain the weight matrix. In the local attention head, the window size is set to w, and each time step only performs attention calculation with the previous and next w time steps. The local attention and global attention are combined into multi-head attention by the following formula: The attention head of multi-head attention is divided into two parts: Local attention head: setting window size , each time step is only related to its predecessor and successor Attention is calculated at each time step.

[0076] Mask matrix construction: For position and location ,if ,but ,otherwise, .

[0077] Local attention calculation:

[0078] In the formula, Represents the query matrix, which is the search part of the attention mechanism and is used to match key-value pairs. represents the key matrix, which is the part being searched in the attention mechanism. is the transpose of this matrix, matching the query matrix, The representative value matrix is ​​the part of the attention mechanism that is actually weighted and summed for subsequent processing. Represents the dimension size of the key, used to scale the dot product attention, represents the mask matrix under local attention, is the attention mechanism under local attention, is a function that converts similarity scores into probability values.

[0079] Global attention heads: allow a specific global position to be evaluated with respect to all other positions.

[0080] Mask matrix construction: For global position , for all positions , , which means that the global position and all positions are calculated for attention.

[0081] Global attention calculation:

[0082] is the attention mechanism under global attention, Represents the mask matrix under global attention.

[0083] Combination of multiple attention heads: local attention heads and The outputs of the global attention heads are concatenated in the feature dimension to form multi-head attention. In this step, the matrix formed by the local attention and the matrix formed by the global attention have the same number of columns. When concatenating the two, they can be directly added.

[0084] Through a linear layer, the concatenated vector is mapped back to the hidden dimension of the model:

[0085] Where; represents the output weight matrix, Represents the global attention or local attention score calculated by the previous single head, represents the matrix concatenation operation, Represents the output of the multi-head attention mechanism, calculates the attention score through multiple attention heads, and concatenates the results and maps them back to the hidden dimension of the model.

[0086] (4) Perform dropout, residual connection and regularization processing, and use feedforward neural network for nonlinear transformation to finally generate output sequence; Dropout is performed on the data flowing through the multi-head attention layer to prevent overfitting, and the Dropout rate is set to 0.2; residual connection and regularization (Add&Norm). The output after Add&Norm is as follows:

[0087] Where: Represented as the concatenated attention score, represents inter-layer regularization, Represents the inter-layer transformation, where the linear transformation is used.

[0088] The feedforward neural network (FFN) is used to perform nonlinear transformation on the input to enhance the complexity and expressiveness of the model. The formula of FFN is shown as follows:

[0089] Where: To perform the output sequence after Add&Norm, are the parameters of the feedforward neural network.

[0090] The data of the Transformer model finally flows through a fully connected layer to produce the final output sequence, as shown in the following formula:

[0091] Where: The output result of nonlinear transformation using feedforward neural network is: are the parameters of the fully connected layer.

[0092] (5) Change the model parameters, train them separately, and finally select the model with the optimal model parameters as the Transformer prediction model; The model dimensions are 64, 32, 16 and 8, the number of encoder layers is 6 and 3, the number of multi-head attention layers is 8 and 4, and the number of epochs is 500, and trial training is performed respectively. The optimal parameters are selected after comprehensively considering the model complexity, prediction accuracy and convergence speed. Of course, those skilled in the art can modify the model parameters to other parameters according to actual conditions.

[0093] (6) Repeat the above steps until the Transformer prediction model of all training sets is obtained; S5. Determine the sample category of the target well, and use the Transformer prediction model corresponding to the sample category to predict its fracture pressure.

[0094] In this embodiment, the following method is used to determine the sample category of the target well: calculate the Euclidean distance between the logging parameters of the target well to be predicted layer segment and each sample in different training sets in S3, and then sum the Euclidean distances between the target well to be predicted layer segment and the samples in each training set. The training set with the smallest sum of Euclidean distances is the sample category of the target well to be predicted layer segment.

[0095] After determining the sample category of the target well, the logging parameters of the target well's predicted layer section are substituted into the Transformer prediction model corresponding to the sample category to predict its fracture pressure.

[0096] In order to further illustrate the accuracy of the model in this embodiment in predicting the burst pressure, the existing technology and its prediction results are also used for comparison. The final results are shown in Table 1 and Figure 4 As shown. Among them, SVR in Table 1 refers to support vector regression, and GPR refers to Gaussian process regression, which belongs to the existing prediction method; Figure 4 It is the result comparison of the predicted value and the measured value of the model of this embodiment.

[0097] Table 1 Prediction accuracy of different models

[0098] From Table 1, it can be seen that the prediction accuracy of the model in this embodiment is better than that of the existing method. Figure 4 It can be seen that the prediction results of the model in this embodiment are relatively close to the measured values, with a small error, and can be used to guide actual production.

[0099] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. An intelligent prediction method for tight conglomerate reservoir fracture pressure based on an improved Transformer architecture, characterized in that: The following steps are involved: S1. Obtain the fracture pressure data of the fractured wells in the same block and the logging parameters related to the fracture pressure, match the fracture pressure and the logging parameters, and pre-process the data to obtain the initial data; S2, based on CAE-SOM self-organizing map neural network, considering the heterogeneity of the reservoir, clustering the initial data to obtain sample categories under different geological conditions; S3. Based on the generative adversarial network, the sample categories under any geological conditions are expanded to obtain the training set; S4. Combining the global attention and local attention mechanisms, the Transformer model is trained using the training set under any geological condition, and finally the Transformer prediction model under any geological condition is obtained; S5. Determine the sample category of the target well, and use the Transformer prediction model corresponding to the sample category to predict its fracture pressure.

2. The method according to claim 1, characterized in that The logging parameters include depth, natural gamma, acoustic time difference, true formation resistivity, density, well diameter, compensated neutron and natural potential.

3. The method according to claim 1, characterized in that S1 includes the following sub-steps: recording the peak pressure during the staged fracturing construction process through the bottom hole pressure change during the fracturing process; using FSI data to identify the fractured fracturing cluster, and matching the peak pressure with the fracturing cluster according to the principle of principal stress size to obtain the true value of the fracture pressure of the fracturing cluster; recording the logging parameters when the fracturing cluster fractures, that is, matching the fracture pressure of the fracturing cluster with the logging parameters; repeating the above operations until the fracture pressure and logging parameters of all fracturing clusters are matched.

4. The method according to claim 1, characterized in that: In S1, the data preprocessing includes the following sub-steps: based on the piecewise regression interpolation method, the data corresponding to each parameter in the acquired LWD parameters are used as a set of fitting data, and fitted into multiple univariate cubic regression curves. When the absolute difference between the acquired value and the fitted value is greater than 5%, the acquired value is regarded as noise and replaced by the fitted value. For the missing values ​​in the data, the average value of the group of data is used to replace them until the abnormal value processing of all data is completed; The processed data are normalized.

5. The method according to claim 1, characterized in that In S2, the following sub-steps are included: extracting effective heterogeneous features from logging parameters based on the autoencoder; Based on the SOM self-organizing map neural network, the effective heterogeneous features are clustered to obtain the sample clustering results under different geological conditions.

6. The method according to claim 1, characterized in that In S3, the upper limit of sample expansion is 2-10 times the original number of samples.

7. The method according to claim 1, characterized in that S4 includes the following sub-steps: Randomly select a training set and transform its data into vectors in a high-dimensional space; adding position coding information to the burst pressure data in the high-level adjacent space; Through Q, K and V, local attention and global attention are calculated respectively and their calculation results are combined into multi-head attention to obtain the weight matrix; among them, in the local attention head, the window size is set to w, and each time step only performs attention calculation with the previous and next w time steps; local attention and global attention are combined into multi-head attention through the following formula: Perform dropout, residual connection and regularization processing, and use feedforward neural network for nonlinear transformation to finally generate output sequence; Change the model parameters, train them separately, and finally select the model with the optimal model parameters as the Transformer prediction model; Repeat the above steps until the Transformer prediction model of all training sets is obtained; Based on the geological conditions of the layer where the target well in this block is located, the fracture pressure can be predicted by substituting the logging while drilling parameters into the Transformer prediction model of the corresponding geological conditions.

8. The method according to claim 1, characterized in that In S5, the method for determining the sample category of the target well includes the following steps: calculating the Euclidean distance between the logging parameters of the target well to be predicted layer segment and each sample in different training sets in S3, and then calculating the sum of the Euclidean distances between the target well to be predicted layer segment and the samples in each training set. The training set with the smallest sum of Euclidean distances is the sample category of the target well to be predicted layer segment.

Images