An intelligent productivity prediction method for high-water-content compact sandstone reservoirs based on an improved convolutional neural network

Abstract

The application discloses a kind of high water cut compact sandstone reservoir productivity intelligent prediction method behavior identification methods based on improved convolutional neural network, the method is carried out feature extraction to well logging curve, constructs more efficient identification network, specific implementation method is as follows: first, establish the well logging data set X suitable for artificial intelligence model training;Then send into the network model BBC proposed in the application for training, by using the expanded CNN and bidirectional recurrent neural network unit (BiGRU-BiLSTM) architecture to generate fusion features, then by combining local and global feature information, finally obtain the prediction result of network model;Finally, use cross-entropy loss, and use stochastic gradient descent to back-propagate the loss of model, optimize the parameters of network model, to achieve global optimization.

Description

Technical Field

[0001] This invention relates to the fields of well logging productivity prediction technology and machine learning technology, and in particular to an intelligent prediction method for well logging productivity using an extended convolutional neural network. Background Technology

[0002] As major oilfields and research institutions deepen their research on oil and gas resources, the development of unconventional gas reservoirs has gradually become a hot topic. Among these, tight sandstone gas reservoirs typically exhibit poor reservoir properties and strong heterogeneity, presenting challenges such as difficulty in identifying gas layers through well logging, quantitative evaluation of reservoir parameters, and quantitative prediction of production capacity. Currently, production capacity prediction methods for tight sandstone gas reservoirs are beginning to integrate with artificial intelligence.

[0003] The fluid identification gas-sensitive factor 1 (IG) and gas-sensitive factor 2 (IRXP) constructed in the literature “Lu Yingzhong. Identification method of tight sandstone gas layer in Hangjinqi based on diameter expansion correction and new gas-sensitive factor [J]. Progress in Geophysics, 2022(001):037” can highlight the contribution of fluid in the Shilijiahan area and improve the consistency rate of gas-water layer identification. The consistency rate of gas layer and gas-water co-layer identification increased from 68% to 92%. The literature “Liu Yiming, Ye Jiaren, Zhang Fei, et al. Fluid identification and evaluation of tight sandstone reservoir in Sha-2 member of Qibei slope belt of Qikou Depression [J]. Earth Science, 2022,47(5):15” effectively distinguishes high-resistivity oil layer from water layer, finely identifies oil layer (including low-resistivity oil layer), oil-water layer, and water layer by applying the overlay map method. The reservoir fluid saturation calculated by the formula method is consistent with the oil test results. The literature “Liu Chengchuan, Chen Jun, Cheng Hongliang. Well logging evaluation technology and its development prospect of tight sandstone gas reservoir in western Sichuan [J]. Journal of Southwest Petroleum University: Natural Science Edition, 2022, 44(3):11” fully utilizes the reservoir fluid property prediction model based on multi-factor intersection and the reservoir fluid identification radar chart of different well types based on multi-factor discrimination to improve the accuracy of reservoir fluid identification; the literature “Bai Yang, Tan Maojin, Xiao Chengwen, et al. Machine logging fluid identification method of tight sandstone gas reservoir dynamic classification committee [J]. Chinese Journal of Geophysics, 2021” uses gate network to divide the input data into multiple subsets, and then uses decision tree, probabilistic neural network, Bayesian classification, BP neural network and nearest neighbor algorithm to train the subsets to obtain multiple sub-models. Finally, the combiner is used to optimize the combination of sub-models to obtain the best fluid identification model. Analysis of the literature “Zhang Haitao, Fang Yuyang, Li Gaoren, et al. Nuclear magnetic resonance relaxation mechanism and fluid identification method of oil-wet tight sandstone [J]. Petroleum Geophysical Exploration, 2020, 59(3):9” reveals that the geometric mean of the nuclear magnetic resonance difference spectrum and the effective porosity difference are sensitive to fluids. The cross plot constructed using these two parameters can effectively distinguish between oil and water layers. The identification of reservoir fluid types was achieved using nuclear magnetic resonance logging data and the constructed plot. The literature “Chen Jun, Xie Runcheng, Liu Chengchuan, et al. Logging fluid identification and quantitative evaluation of Jurassic tight sandstone gas reservoir in Zhongjiang Gas Field [J]. Natural Gas Industry, 2019(S01):6” provides a detailed analysis of different types of reservoir logging fluids. Based on the well response characteristics, a multi-parameter combination method and a BP neural network method based on new fluid identification factors such as the apparent well curvature index and resistivity invasion correction difference ratio are proposed to carry out reservoir fluid identification work. The literature "Liu Dan, Pan Baozhi, et al. GA-SVM fluid identification method based on high resolution array induction logging [J]. Progress in Geophysics, 2017, 32(5):6" introduces the SVM algorithm (GASVM) based on the genetic algorithm GA optimization into well logging interpretation, establishes a nonlinear model for fluid identification, and solves the problem that conventional well logging techniques and interpretation methods cannot accurately identify fluids in tight sandstone reservoirs.

[0004] This invention addresses the complex logging response characteristics of high-water-bearing tight sandstone. Based on logging data, it employs an ensemble learning algorithm based on an improved dilated convolutional neural network (DCNN). This algorithm utilizes a combined architecture of dilated CNN and bidirectional recurrent neural network units (BiGRU-BiLSTM) to generate fused features. By combining local and global feature information with an attention mechanism, the model's predictive performance is improved. By integrating the fused features with the dilated CNN layers, the trained BBC model demonstrates significantly improved classification results and interpretability of feature extraction on the dataset. Summary of the Invention

[0005] The present invention mainly overcomes the shortcomings of the prior art and provides an intelligent prediction method for well logging productivity based on an extended CNN and bidirectional recurrent neural network unit (BiGRU-BiLSTM) architecture.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] 1. A method for intelligent prediction of productivity in high-water-bearing tight sandstone reservoirs based on dilated convolutional neural networks, characterized by the following steps:

[0008] Step 1: Establish a well logging data sample X suitable for training an artificial intelligence model:

[0009] (1) Select a wells from the target stratum of a certain block, where a is a positive integer. Each well contains 7 logging curves: natural gamma, natural potential, compensated neutron, compensated density, sonic transit time, formation resistivity, and flushing zone resistivity, as well as 4 physical property parameters: permeability, porosity, clay content, and water saturation.

[0010] (2) Based on the logging curve and physical property parameters, two composite parameters DT and AK are constructed, and the logging curve, physical property parameters and composite parameters are used as input feature parameters.

[0011]

[0012] Where Δt is the sonic transit time logging value, μm / s; Δt ma Acoustic transit time of the rock skeleton, μm / s; Δt f The acoustic transit time of the fluid, in μm / s; Neutron porosity, %; Hydrogen index of fluid, m 3 / m 3 ;ρ b Compensated density logging value, g / cm³ 3 ;ρ f Fluid density, g / cm³ 3 ;

[0013] (3) The Z-score standardization method is used to make it conform to a normal distribution, wherein the SP curve is locally standardized and other curves are globally standardized;

[0014] (4) Remove the corresponding non-reservoir segments, mudstone interlayers, top and bottom interfaces of reservoir segments, and missing data segments from the selected curves;

[0015] (5) Sample each well section according to a fixed number of sampling points Q, where Q is a positive integer, so that well sections with different thicknesses of reservoirs have different resolutions, and use them as training samples X for the original curve data;

[0016] Step 2: Construct the BBC network model, feed the training sample X into the BBC network model for feature extraction, and obtain the model recognition result Y. The specific steps are as follows:

[0017] (1) Construct a well logging curve productivity prediction model BBC, which consists of BB, miniMSFP, Inception, and

[0018] The dilated convolution is constructed, and the calculation formula is shown below:

[0019] BBC(X)=Inception(miniMSFP(Inception(Conv 3×3 (Conv 1×1 (Conv 7×7 (BB(X)))))))

[0020] In the formula, Conv 1×1 (·) represents a 1×1 convolution, Conv 3×3 (·) represents a 3×3 convolution, Conv 7×7 (·) represents a 7×7 convolution, BB represents a bidirectional recurrent neural network unit layer, and BB consists of a BiGRU(·) network layer and a BiLSTM(·) network layer. The calculation formula is as follows:

[0021] BB(X) = BiGRU(X) × BiLSTM(X)

[0022] miniMSFP is a small multi-scale fusion block. miniMSFP consists of multi-scale average pooling and 1×1 convolution.

[0023] It is composed of products, and the calculation formula is as follows:

[0024]

[0025] In the formula, GAP(·) represents global pooling, and Conv 1×1 (·) represents a convolution with a 1×1 kernel, and N represents the number of scales in the miniMSFP, N=3; APk×k (·) represents average pooling with a kernel of k×k, where k∈{3, 5, 7};

[0026] Inception represents a multi-scale feature extraction block. Inception consists of a 3×3 convolutional Conv... 3×3 (·) and 1×1 convolution Conv 1×1 The (·) component is calculated using the following formula:

[0027] Inception(X) = Concat(Conv 1×1 (X), Conv 3×3 (Conv 1×1 (X)), Conv 3×3 (Conv 1×1 (X)))

[0028] In the formula, Concat(X1, X2, X3) represents arranging the matrix features of features X1, X2, and X3 in order to obtain X. Cat X1, X2, and X3 are n×c×f w×h matrices; X Cat There are n×c×3×f w×h matrices;

[0029] (2) Input the training sample X into the BBC network model;

[0030] (3) The BB layer in the BBC network model extracts features from the training sample X to obtain long-term features X. BB The calculation formula is as follows:

[0031] X BB =BB(X)

[0032] BB(X) = BiGRU(X) × BiLSTM(X)

[0033] In the formula, BB(X) represents long-term feature extraction of training sample X, X represents training sample, BiGRU(X) is the feature vector output by BiGRU network layer, and BiLSTM(X) is the feature vector output by BiLSTM network layer.

[0034] (4) The expanded CNN layer in the BBC network model quickly extracts training samples X to obtain concatenated features. The calculation formula is as follows:

[0035]

[0036] In the formula, X represents the training samples, concatenation represents concatenation, and RV represents the training samples. BiGRU (·) represents the feature output extracted by BiGUR connected to the CNN layer, RVBiSTM (.) represents the feature output extracted by the BiSTM connected to the CNN layer;

[0037] (6) Output of the CNN layer The data is fed into the BBC for feature extraction, and finally the model prediction matrix X is obtained. out The calculation formula is as follows:

[0038]

[0039] Model prediction matrix X out Let X be an n×s matrix, and let X be the model prediction matrix. out The model's recognition result Y is obtained through the maxout function prediction. Y is a vector of length n, representing the recognition result of AFG-NET for n well logging curve samples in the training sample X. The calculation formula is as follows:

[0040]

[0041] In the formula Y n The nth element of the model recognition result Y is represented by Y. n ∈N, Y n ∈[1, n]; x ns Represents the prediction matrix X out n×s elements, x ns ∈(0,1),j n Represents the prediction matrix X out The column number of the largest element in the nth row, j n ∈[1, s]; s represents the number of feature categories of the well logging curve samples in the training sample X;

[0042] Step 3: Use backpropagation to optimize and update the parameters in the BBC network model, and save the network parameter P that achieves the highest recognition accuracy (Acc) in all training rounds. t The specific steps are as follows:

[0043] (1) Backpropagation of error is adopted, and the cross-entropy loss function is used to measure the true result. The distance to the model's recognition result Y is calculated, and the training loss L, L∈(0,+∞), is calculated using the following formula:

[0044]

[0045] In the formula The cross-entropy loss function;

[0046] (2) Through each training session of the training sample X, the stochastic gradient descent function SGD is used to backpropagate the loss value L in the BBC network model, so that the model parameters P of the i-th round of BBC training are...i Randomly varying the model parameters in the direction of the negative gradient to optimize the network, the formula for updating the model parameters using SGD is as follows:

[0047]

[0048] In the formula, P i P represents the model parameters in the i-th round of BBC training. i-1 η represents the model parameters in the (i-1)th round of BBC training, and η represents the optimization step size of SGD, where η∈(0,1). The training loss L represents the model parameters P during the (i-1)th round of BBC training. i-1 Perform differentiation;

[0049] (3) Calculate the current training round e i Accuracy of AFG-NET model in Chinese i Acc i ∈(0,1), and store the training round e. i AFG-NET model P i e i e represents the current i-th training round. i ∈(1, e m ]; e m For the maximum number of training rounds, e m ∈N;

[0050]

[0051] In the formula, TP is the number of positive samples correctly predicted by the model, TN is the number of negative samples correctly predicted by the model, FP is the number of positive samples incorrectly predicted by the model, and FN is the number of negative samples incorrectly predicted by the model.

[0052] (4) Compare the model recognition accuracy Acc in each training round. i And save Acc i The highest model parameter Pi, used as the deployment parameter Pt, is calculated as follows:

[0053]

[0054] Step 4: Input the test set into the trained model to predict reservoir productivity and obtain the prediction results.

[0055] The innovation of this invention is reflected in:

[0056] (1) The deep learning-based model extracts features from well logging curves using a novel bidirectional recurrent neural network based on BiGRU-BiLSTM hybrid hierarchical attention, employing an expanded CNN method to improve model performance. This addresses the problem that arises when traditional expanded convolutional neural network (CNN) models ignore the correlation between context and gradient dispersion.

[0057] (2) The proposed BBC fully utilizes an extended CNN and a Bi-GRU-BiLSTM architecture to generate fused features. By combining local and global feature information, the predictive performance of the model is improved. By combining the fused features with the extended CNN layers, the trained BBC model shows significantly improved classification results and interpretability of feature extraction.

[0058] Beneficial effects:

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] By employing an expanded CNN approach, this approach addresses the issue that arises when traditional expanded convolutional neural network (CNN) models neglect the correlation between context and gradient dispersion. It fully leverages the expanded CNN and the BiGRU-BiLSTM architecture to generate fused features. By combining local and global feature information with attention mechanisms, the model's predictive performance is improved. Attached Figure Description

[0061] Figure 1 This is a structural diagram of the BBC network model proposed in this patent. The BBC network consists of BB, inception block and miniMSFP.

[0062] Figure 2 This is a simplified diagram of the BB layer model structure proposed in this patent. The BB layer consists of a BiGRU network layer and a Bi LSTM network layer.

[0063] Figure 3 The flowchart for the random sampling process based on correlation screening shows that a screening step targeting the correlation within the training subset is added to the random sampling step. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] Example:

[0066] 1. A method for intelligent prediction of productivity in high-water-bearing tight sandstone reservoirs based on dilated convolutional neural networks, characterized by the following steps:

[0067] Step 1: Establish a well logging data sample X suitable for training an artificial intelligence model:

[0068] Forty-two wells were selected from the target strata of a certain block. Each well contained seven logging curves: natural gamma, spontaneous potential, compensated neutron, compensated density, sonic transit time, formation resistivity, and flushed zone resistivity, as well as four physical property parameters: permeability, porosity, clay content, and water saturation.

[0069] Based on well logging curves and physical property parameters, two composite parameters, DT and AK, are constructed, and the well logging curves, physical property parameters, and composite parameters are used together as input feature parameters.

[0070]

[0071] Where Δt is the sonic transit time logging value, μm / s; Δt ma Acoustic transit time of the rock skeleton, μm / s; Δt f The acoustic transit time of the fluid, in μm / s; Neutron porosity, %; Hydrogen index of fluid, m 3 / m 3 ;ρ b Compensated density logging value, g / cm³ 3 ;ρ f Fluid density, g / cm³ 3 ;

[0072] The Z-score standardization method is used to make it conform to a normal distribution, where the SP curve is locally standardized and other curves are globally standardized.

[0073] Remove the corresponding non-reservoir segments, mudstone interlayers, reservoir segment top and bottom interfaces, and missing data segments from the selected curve;

[0074] Each well section was sampled with a fixed number of sampling points Q=30, so that well sections with different thicknesses of reservoirs had different resolutions, which were used as training samples X for the original curve data.

[0075] Step 2: Construct the BBC network model, feed the training sample X into the BBC network model for feature extraction, and obtain the model recognition result Y. The specific steps are as follows:

[0076] (1) Construct a well logging curve productivity prediction model BBC, which consists of BB, miniMSFP, Inception, and

[0077] The dilated convolution is constructed, and the calculation formula is shown below:

[0078] BBC(X)=Inception(miniMSFP(Inception(Conv 3×3 (Conv 1×1 (Conv 7×7 (BB(X)))))))

[0079] In the formula, Conv 1×1 (·) represents a 1×1 convolution, Conv 3×3 (·) represents a 3×3 convolution, Conv 7×7 (·) represents a 7×7 convolution, BB represents a bidirectional recurrent neural network unit layer, and BB consists of a BiGRU(·) network layer and a BiLSTM(·) network layer. The calculation formula is as follows:

[0080] BB(X) = BiGRU(X) × BiLSTM(X)

[0081] miniMSFP is a small multi-scale fusion block. It is composed of multi-scale average pooling and 1×1 convolution, and the calculation formula is as follows:

[0082]

[0083] In the formula, GAP(·) represents global pooling, and Conv 1×1 (·) represents a convolution with a 1×1 kernel, and N represents the number of scales in the mini MSFP, N=3; AP k×k (·) represents average pooling with a kernel of k×k, where k∈{3, 5, 7}; Inception represents a multi-scale feature extraction block, which is formed by a 3×3 convolution Conv 3×3 (·) and 1×1 convolution Conv 1×1 The (·) component is calculated using the following formula:

[0084] Inception(X) = Concat(Conv 1×1 (X), Conv 3×3 (Conv 1×1 (X)), Conv 3×3 (Conv 1×1 (X)))

[0085] In the formula, Concat(X1, X2, X3) represents arranging the matrix features of features X1, X2, and X3 in order to obtain X. Cat X1, X2, and X3 are n×c×f w×h matrices; X Cat There are n×c×3×f w×h matrices:

[0086] (2) Input the training sample X into the BBC network model;

[0087] (3) The BB layer in the BBC network model extracts features from the training sample X to obtain long-term features X. BB The calculation formula is as follows:

[0088] X BB =BB(X)

[0089] BB(X) = BiGRU(X) × BiLSTM(X)

[0090] In the formula, BB(X) represents the long-term feature extraction of training sample X, where X represents the training sample.

[0091] BiGRU(X) is the feature vector output by the BiGRU network layer; BiLSTM(X) is the feature vector output by the BiI and STM network layers.

[0092] (4) The expanded CNN layer in the BBC network model quickly extracts training samples X to obtain concatenated features. The calculation formula is as follows:

[0093]

[0094] In the formula, X represents the training samples, concatenation represents concatenation, and RV represents the training samples. BiGRU (·) represents the feature output extracted by BiGUR connected to the CNN layer, RV BiSTM (·) represents the feature output extracted by the BiSTM connected to the CNN layer;

[0095] (6) Output of the CNN layer The data is fed into the BBC for feature extraction, and finally the model prediction matrix X is obtained. out The calculation formula is as follows:

[0096]

[0097] Model prediction matrix X out Let X be an n×s matrix, and let X be the model prediction matrix. out The model's recognition result Y is obtained through the maxout function prediction. Y is a vector of length n, representing the recognition result of AFG-NET for n well logging curve samples in the training sample X. The calculation formula is as follows:

[0098]

[0099] In the formula Y n The nth element of the model recognition result Y is represented by Y. n ∈N, Y n∈[1, n]; x ns Represents the prediction matrix X out n×s elements, x ns ∈(0,1),j n Represents the prediction matrix X out The column number of the largest element in the nth row, j n ∈[1, s]; s represents the number of feature categories of the well logging curve samples in the training sample X;

[0100] Step 3: Use backpropagation to optimize and update the parameters in the BBC network model, and save the network parameter P that achieves the highest recognition accuracy (Acc) in all training rounds. t The specific steps are as follows:

[0101] (1) Backpropagation of error is adopted, and the cross-entropy loss function is used to measure the true result. The distance to the model's recognition result Y is calculated, and the training loss L, L∈(0,+∞), is calculated using the following formula:

[0102]

[0103] In the formula The cross-entropy loss function;

[0104] (2) Through each training session of the training sample X, the stochastic gradient descent function SGD is used to backpropagate the loss value L in the BBC network model, so that the model parameters P of the i-th round of BBC training are... i Randomly varying the model parameters in the direction of the negative gradient to optimize the network, the formula for updating the model parameters using SGD is as follows:

[0105]

[0106] In the formula, P i P represents the model parameters in the i-th round of BBC training. i-1 η represents the model parameters in the (i-1)th round of BBC training, and η represents the optimization step size of SGD, where η∈(0,1). The training loss L represents the model parameters P during the (i-1)th round of BBC training. i-1 Perform differentiation;

[0107] (3) Calculate the current training round e i Accuracy of AFG-NET model in Chinese i Acc i ∈(0,1), and store the training round e. i AFG-NET model P i e i e represents the current i-th training round.i ∈(1, e m ]; e m For the maximum number of training rounds, e m ∈N;

[0108]

[0109] In the formula, TP is the number of positive samples correctly predicted by the model, TN is the number of negative samples correctly predicted by the model, FP is the number of positive samples incorrectly predicted by the model, and FN is the number of negative samples incorrectly predicted by the model.

[0110] (4) Compare the model recognition accuracy Acc in each training round. i And save Acc i The highest model parameter Pi, used as the deployment parameter Pt, is calculated as follows:

[0111]

[0112] Step 4: Input the test set into the trained model to predict reservoir productivity and obtain the prediction results.

Claims

1. A method for intelligent prediction of productivity in high-water-bearing tight sandstone reservoirs based on an improved convolutional neural network, characterized in that... Includes the following steps: Step 1: Establish a well logging data sample X suitable for training an artificial intelligence model: (1) Select a wells from the target stratum of a certain block, where a is a positive integer. Each well contains 7 logging curves: natural gamma, natural potential, compensated neutron, compensated density, sonic transit time, formation resistivity, and flushing zone resistivity, as well as 4 physical property parameters: permeability, porosity, clay content, and water saturation. (2) Based on the logging curve and physical property parameters, two composite parameters DT and AK are constructed, and the logging curve, physical property parameters and composite parameters are used as input feature parameters. Where Δt is the sonic transit time logging value, μm / s; Δt ma Acoustic transit time of the rock skeleton, μm / s; Δt f The acoustic transit time of the fluid, in μm / s; Neutron porosity, %; Hydrogen index of fluid, m 3 / m 3 ;ρ b Compensated density logging value, g / cm³ 3 ;ρ f Fluid density, g / cm³ 3 ; (3) The Z-score standardization method is used to make it conform to a normal distribution, wherein the SP curve is locally standardized and other curves are globally standardized; (4) Remove the corresponding non-reservoir segments, mudstone interlayers, top and bottom interfaces of reservoir segments, and missing data segments from the selected curves; (5) Sample each well section according to a fixed number of sampling points Q, where Q is a positive integer, so that well sections with different thicknesses of reservoirs have different resolutions, and use them as training samples X for the original curve data; Step 2: Construct the BBC network model, feed the training sample X into the BBC network model for feature extraction, and obtain the model recognition result Y. The specific steps are as follows: (1) Construct a well logging curve productivity prediction model BBC. BBC is composed of BB, miniMSFP, Inception and void convolution. The calculation formula is as follows: BBC(X)=Inception(miniMSFP(Inception(Conv 3×3 (Conv 1×1 (Conv 7×7 (BB(X))))))) In the formula, Conv 1×1 (·) represents a 1×1 convolution, Conv 3×3 (·) represents a 3×3 convolution, Conv 7×7 (·) represents a 7×7 convolution, BB represents a bidirectional recurrent neural network unit layer, and BB consists of a BiGRU(·) network layer and a BiLSTM(·) network layer. The calculation formula is as follows: BB(X) = BiGRU(X) × BiLSTM(X) miniMSFP is a small multi-scale fusion block. It is composed of multi-scale average pooling and 1×1 convolution, and the calculation formula is as follows: In the formula, GAP(·) represents global pooling, and Conv 1×1 (·) represents a convolution with a 1×1 kernel, and N represents the number of scales in the mini MSFP, N=3; AP k×k (·) represents average pooling with a kernel of k×k, where k∈{3, 5, 7}; Inception represents a multi-scale feature extraction block, which is formed by a 3×3 convolution Conv 3×3 (·) and 1×1 convolution Conv 1×1 The (·) component is calculated using the following formula: Inception(X)=Concat(Conv 1×1 (X),Conv 3×3 (Conv 1×1 (X)),Conv 3×3 (Conv 1×1 (X))) In the formula, Concat(X1, X2, X3) represents arranging the matrix features of features X1, X2, and X3 in order to obtain X. Cat X1, X2, and X3 are n×c×f w×h matrices; X Cat There are n×c×3×f w×h matrices; (2) Input the training sample X into the BBC network model; (3) The BB layer in the BBC network model extracts features from the training sample X to obtain long-term features X. BB The calculation formula is as follows: X BB =BB(X) BB(X) = BiGRU(X) × BiLSTM(X) In the formula, BB(X) represents long-term feature extraction of training sample X, X represents training sample, BiGRU(X) is the feature vector output by BiGRU network layer, and BiLSTM(X) is the feature vector output by BiLSTM network layer. (4) The expanded CNN layer in the BBC network model quickly extracts training samples X to obtain concatenated features. The calculation formula is as follows: In the formula, X represents the training sample, concatenation represents concatenation, and RV BiGRU (·) represents the feature output extracted by BiGUR connected to the CNN layer, RV BiSTM (·) represents the feature output extracted by BiSTM connected to the CNN layer; (6) Output of the CNN layer The data is fed into the BBC for feature extraction, and finally the model prediction matrix X is obtained. out The calculation formula is as follows: Model prediction matrix X out Let X be an n×s matrix, and let X be the model prediction matrix. out The model's recognition result Y is obtained through the maxout function prediction. Y is a vector of length n, representing the recognition result of AFG-NET for n well logging curve samples in the training sample X. The calculation formula is as follows: In the formula Y n The nth element of the model recognition result Y is represented by Y. n ∈N, Y n ∈[1, n]; x ns Represents the prediction matrix X out n×s elements, x ns ∈(0,1),j n Represents the prediction matrix X out The column number of the largest element in the nth row, j n ∈[1, s]; s represents the number of feature categories of the well logging curve samples in the training sample X; Step 3: Use backpropagation to optimize and update the parameters in the BBC network model, and save the network parameter P that achieves the highest recognition accuracy (Acc) in all training rounds. t The specific steps are as follows: (1) Backpropagation of error is adopted, and the cross-entropy loss function is used to measure the true result. The distance to the model's recognition result Y is calculated, and the training loss L, L∈(0,+∞), is calculated using the following formula: In the formula The cross-entropy loss function; (2) Through each training session of the training sample X, the stochastic gradient descent function SGD is used to backpropagate the loss value L in the BBC network model, so that the model parameters P of the i-th round of BBC training are... i Randomly varying the model parameters in the direction of the negative gradient to optimize the network, the formula for updating the model parameters using SGD is as follows: In the formula, P i P represents the model parameters in the i-th round of BBC training. i-1 η represents the model parameters in the (i-1)th round of BBC training, and η represents the optimization step size of SGD, where η∈(0,1). The training loss L represents the model parameters P during the (i-1)th round of BBC training. i-1 Perform differentiation; (3) Calculate the current training round e i Accuracy of AFG-NET model in Chinese i Acc i ∈(0,1), and store the training round e. i AFG-NET model P i e i e represents the current i-th training round. i ∈(1, e m ]; e m For the maximum number of training rounds, e m ∈N; In the formula, TP is the number of positive samples correctly predicted by the model, TN is the number of negative samples correctly predicted by the model, FP is the number of positive samples incorrectly predicted by the model, and FN is the number of negative samples incorrectly predicted by the model. (4) Compare the model recognition accuracy Acc in each training round. i And save Acc i The highest model parameter Pi, used as the deployment parameter Pt, is calculated as follows: Step 4: Input the test set into the trained model to predict reservoir productivity and obtain the prediction results.

Images