Method for predicting shear wave velocity and density logging data of shale formation cycle characteristics

Abstract

The invention discloses a method for predicting shear wave velocity and density logging data of shale formation cycle characteristics, and relates to the technical field of seismic exploration and development. According to the method, input data is divided into a plurality of time sequence fragments, target values are extracted as labels, and the problem of prediction precision reduction caused by a non-unique mapping relation between samples and labels is solved. Through combination of a space-time attention mechanism and a CNN-GRU hybrid model, an STACGN fusion network is constructed, and the capture capability of the model on waveform sample structure features is effectively enhanced. The network can more accurately capture the spatial-temporal characteristics of the input waveform through training and prediction, so that the precision of shear wave speed and density prediction is remarkably improved. Compared with a traditional method, the method has the advantages that the space-time structure characteristics of data can be fully utilized, the prediction result is optimized, and a more accurate measurement basis is provided for shale formation exploration.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic exploration and development, and in particular to a method for predicting shear wave velocity and density logging data of shale formation cycle characteristics. Background Art

[0002] P-wave velocity, S-wave velocity, and rock density play important roles in AVA and AVO analysis of seismic data, prestack seismic inversion, reservoir property prediction, and fluid identification. These three types of data are particularly crucial for seismic inversion based on well logging data. However, in actual well logging, S-wave velocity data is relatively scarce due to the difficulty in interpreting it and the high cost of obtaining it. Therefore, accurately predicting S-wave velocity is particularly important. Furthermore, wellbore expansion and contraction can significantly impact density logging, causing distortion, loss, or even loss of data in some well sections, complicating subsequent comprehensive interpretation. Therefore, accurately predicting S-wave velocity and density is crucial.

[0003] Traditional shear wave velocity and density predictions are mainly divided into three types: empirical model method, multivariate fitting method and rock physics modeling method. Chinese patent application number CN202211504989.0 applies for a method, device and application for predicting shear wave velocity of shale formations, and Chinese patent application number CN202310975346.2 applies for a method for predicting shear wave time difference in continental shale gas reservoirs. Both mentioned the empirical model method and multivariate fitting method for fitting the shear wave velocity of shale formations: the first article focuses on the organic matter composition, which affects both the longitudinal and shear wave velocities. The intersection diagram technology is used to fit the longitudinal wave velocity and the shear wave velocity represented by the organic matter content. The resulting multivariate fitting formula is used to predict wells with missing shear waves. The second article focuses on the influence of clay mineral composition and porosity on longitudinal and shear wave velocities, and uses the multivariate fitting method to fit the relationship between shear wave velocity and natural gamma and neutron porosity. Both inventions use the multivariate fitting method to derive the relationship between shear wave velocity and other conventional logging curves. Compared with other methods, this method is simple to calculate, but neither of these two patents fully considers the factors affecting shear wave velocity, such as mineral content, porosity, organic matter content, etc. In addition, the multivariate fitting method lacks detailed information on the shear wave velocity, and research on nonlinear fitting methods affected by multiple factors should be strengthened.

[0004] Chinese patent application number CN202011201370.3 applies for a method, storage medium, and system for predicting shear wave velocity in organic-rich rocks. The patent mentions a rock physics model approach for predicting shear wave velocity in shale formations: a well logging interpretation method interprets logging data to obtain organic matter content, mineral content, porosity, and saturation, equating kerogen in organic-rich rocks to a mineral matrix. A shale rock physics model is established that relates the rock's longitudinal and shear wave velocities to kerogen content, porosity, matrix minerals, critical porosity, and consolidation coefficient. This patent provides a rock physics model for predicting shear wave velocity in organic-rich rocks. This model can improve the accuracy of velocity prediction for organic-rich rocks when logging data is complete and interpretations are reliable. However, the patent requires input of multiple shale parameters, such as mineral composition, porosity, and organic matter content. These parameters are not available in some older wells, limiting its application. Furthermore, the complex pore structure of shale formations can reduce the accuracy of the shale rock physics model.

[0005] Chinese patent application number CN202111296927.0 discloses a deep learning-based shear wave velocity prediction method. The method uses a deep feedforward neural network (DFNN) algorithm to establish a nonlinear relationship between the compressional wave velocity curve (Vp), the neutron porosity curve (CNL), the density curve (DEN), the well logging interpretation data porosity curve (POR), the shale content curve (SH), and the shear wave velocity curve (Vs), thereby constructing a shear wave velocity prediction model. The patent provides a deep learning-based shear wave velocity prediction method that replaces empirical formulas and rock physics modeling methods to improve the accuracy and efficiency of shear wave velocity estimation. However, the patent does not explain the impact of the data processing characteristics of the deep learning network on the prediction results.

[0006] The Chinese patent application with application number CN202111241541.X has applied for a density prediction method, apparatus, device and storage medium based on a neural network, and mentioned the deep learning method for predicting density: using a deep learning network to explore the intrinsic relationship between rock physical parameters, establish a nonlinear relationship model between P-wave and S-wave velocity and density, and obtain a density prediction model based on a deep learning network. This patent provides a density prediction method that replaces the Gardner empirical formula method deep learning network, fully exploring the nonlinear relationship between P-wave and S-wave velocity and density. However, this patent does not process the input samples of the deep learning network model, and essentially does not change the single-point to single-point mapping relationship, that is, it does not change the relationship between the same input and different output. Summary of the Invention

[0007] In order to solve the problems of low prediction accuracy of traditional empirical models and multivariate fitting methods and the technical problem that rock physics modeling methods are difficult to reconstruct multiple curves simultaneously, the present invention discloses a method for predicting shear wave velocity and density logging data of shale formation cyclic characteristics. It overcomes the problem of decreased prediction accuracy caused by the non-unique mapping relationship between samples and labels, and improves the generalization ability and prediction accuracy of the model; a deep learning fusion network model combining convolutional neural network (CNN), gated recurrent unit (GRU) and attention mechanism is established to better capture the structural characteristics of input waveform samples, while improving the shear wave velocity and density accuracy.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] The method for predicting the shear wave velocity and density logging data of the shale formation cyclic characteristics includes the following steps:

[0010] (1) Use the StandardScaler function to normalize and standardize the six types of logging data so that the processed logging data conform to the standard normal distribution;

[0011] (2) Divide the input data into several time series segments as model features through a sliding window method, extract the target value from the corresponding intermediate position as the label, and divide it into training set and test set;

[0012] (3) Combining the convolutional neural network (CNN) with the gated recurrent neural unit (GRU) to build a CNN-GRU hybrid model, and integrating the spatiotemporal attention mechanism with the hybrid neural network to form a STACGN fusion network;

[0013] (4) Inputting the conventional logging curve of the training well after processing in step (1) into the STACGN fusion network for training;

[0014] (5) Inputting the data preprocessed in step (1) and the test set divided and processed in step (2) into the fusion network for training and prediction;

[0015] (6) Using R 2 As the evaluation index of STACGN fusion network model, network evaluation is carried out.

[0016] Furthermore, the conversion function is as follows: in step (1), the six types of logging data include natural gamma curve, compressional wave velocity curve, neutron porosity curve, natural potential curve, shear wave velocity curve and density curve;

[0017] The conversion function used in preprocessing is:

[0018]

[0019] Among them, X i is the well logging data, X μ and X σ are the mean and variance of the logging data, Y i The data are standardized.

[0020] Furthermore, in step (3), in the STACGN fusion network,

[0021] [X 1,1 ...X t,m ] represents two-dimensional sample data with t rows and m columns;

[0022] [α1...α m ] is the weight vector of the spatial attention layer;

[0023] α m is the weight of the mth spatial feature;

[0024] [T 1,1 ...T t,m ] is the feature extracted from the 2DCNN layer;

[0025] T t,m is the feature value extracted by the mth convolutional layer;

[0026] [T 1,1 α1...T t,m α m ] represents the output of the spatial attention layer;

[0027] h t,n1 , H t,n2 The hidden states of the first and second layer GRUs at time step t have n1 features and n2 features respectively;

[0028] [β1...β t ] is the weight vector of the temporal attention layer;

[0029] β t is the weight of the feature at the tth time step;

[0030] [H' 1,1 β1...H' t,n2 β t ] is the output of the temporal attention layer.

[0031] Furthermore, in step (4), conventional logging curves include natural gamma curves, compressional wave velocity curves, neutron porosity curves and natural potential curves.

[0032] Furthermore, in step (4), the STACGN fusion network combines convolutional neural networks and gated recurrent neural units, uses convolutional layers to extract spatial features of logging data, and captures temporal dependencies of sequence data through GRU layers. During model training, mean square error (MSE) is used as the loss function, and Adam optimizer is used for parameter update. During training, the Dropout mechanism is introduced, and the learning rate is adjusted. The performance of the model is evaluated through cross-validation, and a lower error is obtained on the training set. The loss function is as follows:

[0033]

[0034] Where MSE is the mean square error, n is the total number of samples, and Y i is the true value of the i-th sample in the dataset, is the predicted value for this sample.

[0035] Furthermore, in step (5), R 2 The coefficient of determination is used to measure the degree of fit between the model's predicted value and the actual value. The closer its value is to 1, the stronger the model's ability to explain the data. The expression is as follows:

[0036]

[0037] Among them, y i is the true label, is the mean of the true values, is the predicted value of the network, and n is the number of samples.

[0038] The beneficial effects of the present invention are as follows: compared with the existing shear wave velocity and density logging curve prediction technology, the present invention has the following advantages:

[0039] (1) Drawing on the geological idea of ​​dividing small layers by comparing waveform results, this method uses a sliding time window technique to expand the dimension of single-point samples into waveform structure samples, achieving a high-dimensional space representation of single-point labels. This overcomes the problem of reduced prediction accuracy caused by the non-unique mapping relationship between samples and labels, and improves the model's generalization ability and prediction accuracy. Compared with conventional fixed segmentation methods, this method is more flexible and efficient, and can better improve the model's learning effect.

[0040] (2) Based on the different characteristics of neural network data processing, the gated recurrent network (GRU) is selected to extract the autocorrelation features of the logging curve, and the CNN is used to extract the cross-correlation features of different logging curves. The attention mechanism improves the weight coefficient of the deep learning network of the logging curve with high correlation with shear wave velocity and density, forming a convolution (CNN) + gated recurrent (GRU) fusion network method for predicting shear wave velocity and density that integrates the spatiotemporal attention mechanism (ATTENTION). The weight distribution of the spatiotemporal attention layer of the fusion deep learning network is consistent with the size law of the autocorrelation and cross-correlation of conventional logging curves. The established deep learning network fusion method can effectively improve the generalization ability of conventional shear wave velocity and density curve prediction, taking into account both generalization ability and prediction accuracy.

[0041] (3) The existing empirical model method and multivariate fitting method have strong generalization ability but low accuracy. The rock physics modeling method has high prediction accuracy and generalization ability, but cannot predict the shear wave velocity and density curves at the same time. Combined with examples, a comparative analysis is conducted from the aspects of deep learning network structure, generalization ability, traditional methods and sample set division, which shows that the proposed network has higher prediction accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a structural diagram of the STACGN fusion network in the present invention;

[0043] Figure 2 VS, DEN training and prediction flow chart of the STACGN fusion network in an embodiment of the present invention;

[0044] Figure 3 This is a cross-plot of VS, DEN and other logging data in an embodiment of the present invention;

[0045] Figure 4 is the autocorrelation coefficient of the well logging data in the embodiment of the present invention;

[0046] Figure 5 This is a sample partition diagram for multiple pairs in an embodiment of the present invention;

[0047] Figure 6 The temporal feature weight distribution of well logging data in an embodiment of the present invention (with or without temporal attention layer) (left figure) VS; (right figure) DEN;

[0048] Figure 7 The spatial feature weight distribution of well logging data in an embodiment of the present invention (with or without spatial attention layer) (left figure) VS; (right figure) DEN;

[0049] Figure 8 A comparison chart of the predicted results of shear wave velocities of different networks in an embodiment of the present invention;

[0050] Figure 9 A comparison chart of prediction results of different network densities in an embodiment of the present invention;

[0051] Figure 10 A comparison diagram of the predicted results of shear wave velocities under different time windows in an embodiment of the present invention;

[0052] Figure 11 A comparison diagram of density prediction results under different windows at different times in an embodiment of the present invention;

[0053] Figure 12 This is a comparison chart of the prediction results of dy-2 shear wave velocity in an embodiment of the present invention;

[0054] Figure 13 4 is a comparison chart of the prediction results of dy-2 density in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0056] Aiming to solve the problems of low prediction accuracy of traditional empirical models and multivariate fitting methods and the difficulty of rock physics modeling methods in simultaneously reconstructing multiple curves, this method aims to better extract important features in both spatial and depth directions at the same time. Drawing on the idea of ​​small-layer comparison in geology, this method converts the single-point sample structure into a waveform sample structure and establishes a deep learning fusion network model that combines CNN, gated recurrent unit (GRU) and attention mechanism to better capture the structural characteristics of the input waveform samples while improving the accuracy of shear wave velocity and density.

[0057] The present invention discloses a method for predicting shear wave velocity and density logging data of shale formation cyclic characteristics, comprising the following steps:

[0058] (1) Use the StandardScaler function to normalize and standardize the logging data so that the processed logging data conforms to the standard normal distribution, which is conducive to model training.

[0059] The conversion function is as follows: In step (1), the conversion function used in preprocessing is:

[0060]

[0061] Among them, Xi is the well logging data, X μ and X σ are the mean and variance of the logging data, Y i The data are standardized.

[0062] (2) To achieve accurate prediction of shear wave velocity and density, the input data is divided into several time series segments as model features using a sliding window approach, and target values ​​are extracted from the corresponding intermediate positions as labels. This approach ensures that the model can learn the correlation between past observations and establish a mapping relationship with the target labels, ensuring that the model can learn the temporal dependency between the input sequence and the target sequence during training. At the same time, by extracting overlapping sequence segments from different positions, the number of training samples is increased, enhancing the generalization ability of the model. This construction method provides efficient and reasonable input-output pairs for sequence prediction tasks, thereby improving the accuracy of predictions.

[0063] (3) In order to achieve accurate prediction of the target logging curve, the present invention innovatively introduces deep learning technology, aiming to use the powerful expression ability of neural networks to directly learn the complex mapping relationship between input and output, optimize features and improve the learning performance of subsequent models. The CNN part is responsible for extracting spatial features from logging data, the GRU is responsible for capturing the temporal dependencies of logging data, and the spatiotemporal attention mechanism focuses on extracting important spatial and temporal features of logging data. By integrating and extracting important spatial and temporal domain information, the generalization ability and recognition accuracy of the model are significantly enhanced. A CNN-GRU hybrid model is constructed by combining convolutional neural network (CNN) and gated recurrent neural unit (GRU), and the spatiotemporal attention mechanism is fused with the hybrid neural network to form a STACGN fusion network ( Figure 1 shown).

[0064] In the STACGN fusion network, [X 1,1 ...X t,m ] represents two-dimensional sample data with t rows and m columns; [α1...α m ] is the weight vector of the spatial attention layer; α m is the weight of the mth spatial feature; [T 1,1 ...T t,m ] is the feature extracted from the 2DCNN layer; T t,m is the feature value extracted by the mth convolutional layer; [T 1,1 α1...T t,m α m ] represents the output of the spatial attention layer; h t,n1 , H t,n2The hidden states of the first and second layer GRUs with n1 and n2 features at the tth time step respectively; [β1...β t ] is the weight vector of the temporal attention layer; β t is the weight of the feature at the tth time step; [H' 1,1 β1...H' t,n2 β t ] is the output of the temporal attention layer.

[0065] (4) The conventional logging curves of the training wells after previous processing are input into the STACGN fusion network for training.

[0066] Conventional logging curves include natural gamma curves, compressional wave velocity curves, neutron porosity curves and natural potential curves.

[0067] The STACGN fusion network combines convolutional neural networks and gated recurrent neural units, using convolutional layers to extract spatial features of logging data and GRU layers to capture the temporal dependencies of sequence data. During model training, the mean square error (MSE) is used as the loss function, and the Adam optimizer is used for parameter updates. The Dropout mechanism is also introduced during training, and the learning rate is adjusted. The performance of the model is evaluated through cross-validation, and a low error is obtained on the training set, indicating that the model has good generalization ability in predicting shear wave velocity and density.

[0068] The loss function is as follows:

[0069]

[0070] Where MSE is the mean square error, n is the total number of samples, and Y i is the true value of the i-th sample in the dataset, is the predicted value for this sample.

[0071] The MSE formula can be used to evaluate the quality of the regression model because it measures the error between the true value and the predicted value and sums and averages the squares. If the MSE is small, it means that our model performs more accurately in prediction; conversely, if the MSE is large, it means that the model performs poorly and the gap between the true value and the predicted value is large.

[0072] (5) The test set that has undergone the previous data preprocessing and data sample set division is input into the fusion network for training and prediction.

[0073] (6) Using R 2 As the evaluation index of the fusion network model, network evaluation is carried out.

[0074] R 2The coefficient of determination is used to measure the degree of fit between the model's predicted value and the actual value. The closer its value is to 1, the stronger the model's ability to explain the data. The expression is as follows:

[0075]

[0076] Among them, y i is the true label, is the mean of the true values, is the predicted value of the network, and n is the number of samples.

[0077] By R 2 To evaluate the generalization ability of the model on different data sets, it shows that the fusion network model can still maintain high performance on unseen data.

[0078] Example

[0079] In order to verify the effectiveness of this method, the shale reservoirs of wells dy-1 and dy-2 in a certain work area of ​​Dongying Depression were selected as the research objects. The main process ( Figure 2 ) includes the following parts:

[0080] (1) Correlation analysis

[0081] The data sets are all from the same target interval, which can ensure the similarity of the data sets in the horizontal direction. The characteristic curves input in the sample set are natural gamma (GR), compressional wave velocity (VP), neutron porosity (CNL), and formation resistivity (RT), and the label set is shear wave velocity (VS) and density (DEN). The cross-plot of the target logging curve and other logging curves is used to analyze the correlation between them ( Figure 3 ), among which the correlations between VS and other logging data from high to low are: compressional wave velocity (VP), neutron porosity (CNL), resistivity (RT), and gamma (GR), with determination coefficients of 0.925, 0.668, 0.428, and 0.325, respectively.

[0082] The correlations of DEN from high to low are: compressional velocity (VP), neutron porosity (CNL), gamma (GR), and resistivity (RT). The coefficients of determination are 0.695, 0.658, 0.342, and 0.215, respectively. In addition, the autocorrelation function (ACF) is used to analyze the autocorrelation of conventional logging data ( Figure 4 ), its autocorrelation decreases with the increase of lag distance, which shows that the underground sedimentary strata are continuously changing.

[0083] (2) Sample dimension removal preprocessing

[0084] The StandardScaler function is used to normalize and standardize the logging data so that the processed logging data conforms to the standard normal distribution, removing the instability of the deep learning network model caused by the dimensional difference of the logging curve, which is beneficial to the training of the model.

[0085] (3) Sliding time window technology expands single-point samples into waveform structure samples

[0086] In the time series direction, the sliding window technology is used to process the single point sample into a waveform structure sample considering the cyclic characteristics of the shale formation, forming a many-to-one method to divide the sample set ( Figure 5 ), this method fully ensures the autocorrelation characteristics of the logging data and reflects the cyclic characteristics of the underground sedimentary shale formation.

[0087] (4) Analysis of the weight of the attention layer of deep learning networks

[0088] To verify the effectiveness of the attention mechanism in enhancing the sensitivity of deep learning networks to key spatiotemporal features, two network structures were constructed: with and without the spatiotemporal attention mechanism. Figure 6 The weight distribution of the well logging data when the temporal attention layer is included or not is shown. When the target label is shear wave velocity or density, the weight distribution of the spatial attention layer is VP, CNL, RT, GR (from high to low). Figure 7 Left) and VP, CNL, GR, RT ( Figure 7 Right), this distribution is consistent with Figure 3 The cross-correlation between the target curve and other curves is consistent. However, in the network without the attention mechanism, the weight distribution does not have this regularity. Figure 7 The weight distribution of the time series attention layer is shown in the figure. 15 sampling points in the conventional well logging data are selected as samples, and VS or DEN in the middle of the sample is used as the label. When predicting VS or DEN, the time series attention layer gives the highest weight ratio to the sample data at the label position, and as the distance from the label position increases, the attention weight generally decreases, which is consistent with the Figure 3 The autocorrelation distribution law of conventional logging data is consistent with that of the conventional logging data.

[0089] (5) Comparison of prediction effects of different deep learning networks

[0090] This paper analyzes the prediction results of STACGN, a two-dimensional convolutional neural network, a gated recurrent unit, and multivariate fitting. The network structures are shown in Table 1. All networks use Adam with momentum as the optimization algorithm, which combines the advantages of AdaGrad and RMSProp algorithms and has strong advantages in processing large-scale data, parameter optimization, and non-stationary objectives. At the same time, a dropout layer is added to randomly discard neurons to increase the generalization of the network.

[0091] Table 1. STACGN, 2D convolutional neural network, and gated recurrent unit network structures

[0092]

[0093] The present invention compares the performance of different deep learning network reconstructions. The dy-1 well logging data is selected to train the two-dimensional convolutional neural network, the gated recurrent unit and the STACGN network. Overall, the prediction results of the proposed network VS and DEN are better than those of the other two networks ( Figure 8-Figure 9 ), and compared to the prediction results of multivariate fitting, the prediction results of all three networks were significantly improved. When reconstructing shear wave logging curves, the correlation coefficients of the STACGN fusion network, CNN, GRU, and multivariate fitting methods were 0.91, 0.84, 0.83, and 0.66, respectively. The correlation coefficients for density logging curve reconstruction methods were 0.89, 0.81, 0.79, and 0.55, respectively. The correlation coefficients also show that the STACGN fusion network has the best performance in reconstructing shear wave and density logging curves, while multivariate fitting has the worst performance. This demonstrates the clear advantage of deep learning networks in capturing the complex nonlinear relationships between the target curve and other curves.

[0094] (6) Comparison of prediction effects with different sliding window lengths

[0095] To improve the generalization ability and prediction accuracy of the model, this paper compares and analyzes the prediction results of the dy-1 well under different sliding window parameters. By adjusting the sliding window length, its impact on model performance is studied. It is found that when the sliding window parameter is 15, it can better capture local features and help the model learn longer time dependencies.

[0096] In order to further verify the prediction performance of the proposed network, R 2 To conduct a more accurate quantitative evaluation of the prediction results, R 2 The larger the value is, the stronger the correlation between the predicted data and the label data is. 2 , ( Figure 10-11 ) When the window length is 5, the density and shear wave velocity R 2 are 0.77 and 0.79 respectively; when the window length is 15, the R 2 are 0.91 and 0.93 respectively; when the window length is 25, the R 2 Further analysis shows that when the window length is 15, a balance can be achieved between local details and global trends, thereby improving the prediction accuracy of the model on the test set.

[0097] (7) Analysis of blind well data prediction results

[0098] To verify the generalization capability of the invention, the proposed method was used to predict the missing shear wave velocity and density logs for the dy-2 well. By using partial logging data from the dy-2 well as input and utilizing the preferred sliding window parameters of the invention, the results showed that the model was able to accurately reconstruct the missing shear wave velocity and density curves. Figure 12-13 ) Compared with the actual measurement data, the determination coefficient (R 2 ) shows that the proposed method has a high prediction accuracy, and the coefficient of determination of shear wave velocity and density (R 2 ) reached 0.87 and 0.86, respectively. The coefficients of determination for the shear wave velocity and density predictions using the 2D convolutional neural network method were 0.79 and 0.78. The coefficients of determination for the shear wave velocity and density predictions using the gated recurrent unit network method were 0.78 and 0.76, respectively. This further demonstrates the good generalization and robustness of this method when processing data from other wells.

[0099] In order to solve the problems of shear wave velocity data shortage caused by the difficulty in interpreting shear wave velocity and the high acquisition cost in actual well logging, as well as the distortion or loss of density logging data caused by problems such as wellbore collapse and instrument failure, the present invention adopts a deep learning method integrating spatiotemporal attention mechanism to predict shear wave velocity and density logging curves.

[0100] First, drawing on the geological concept of dividing small layers by comparing waveform results, a sliding time window technique was used to expand single-point samples into waveform structure samples, achieving a high-dimensional spatial representation of single-point labels and overcoming the problem of reduced prediction accuracy caused by non-unique mappings between samples and labels. Then, based on the different characteristics of neural network data processing, a gated recurrent network (GRU) was used to extract the autocorrelation features of well logging curves, and a CNN was used to extract the cross-correlation features of different well logging curves. An attention mechanism was used to increase the weight coefficients of the deep learning network for well logging curves with high correlation with shear wave velocity and density, resulting in a convolutional (CNN) + gated recurrent (GRU) fusion network method for predicting shear wave velocity and density that incorporates a spatiotemporal attention mechanism (ATTENTION). Measured shale formation data was then used to interpret the spatiotemporal attention layer weight distribution of the fused deep learning network, which is consistent with the autocorrelation and cross-correlation patterns of conventional well logging curves. This demonstrates that the deep learning network developed in this method, integrated with conventional shear wave velocity and density prediction methods, has strong generalization capabilities, balancing generalization and prediction accuracy. Finally, the proposed network has higher prediction accuracy and generalization ability through multiple analysis and comparison of different deep learning network structures and generalization capabilities, traditional methods, and sample set division time windows.

[0101] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for predicting cyclic characteristics of shale formations using shear wave velocity and density logging data, characterized in that: The following steps are involved: (1) Use the StandardScaler function to normalize and standardize the six types of logging data so that the processed logging data conform to the standard normal distribution; (2) Divide the input data into several time series segments as model features through a sliding window method, extract the target value from the corresponding intermediate position as the label, and divide it into training set and test set; (3) Combining the convolutional neural network (CNN) with the gated recurrent neural unit (GRU) to build a CNN-GRU hybrid model, and integrating the spatiotemporal attention mechanism with the hybrid neural network to form a STACGN fusion network; (4) Inputting the conventional logging curve of the training well after processing in step (1) into the STACGN fusion network for training; (5) Inputting the data preprocessed in step (1) and the test set divided and processed in step (2) into the fusion network for training and prediction; (6) Using R 2 As the evaluation index of STACGN fusion network model, network evaluation is carried out.

2. The method for predicting shale formation cyclic characteristics based on shear wave velocity and density logging data according to claim 1, wherein: The conversion function is as follows: In step (1), the six types of logging data include natural gamma curve, compressional wave velocity curve, neutron porosity curve, natural potential curve, shear wave velocity curve and density curve; The conversion function used in preprocessing is: Among them, X i is the well logging data, X μ and X σ are the mean and variance of the logging data, Y i The data are standardized.

3. The method for predicting shale formation cyclic characteristics based on shear wave velocity and density logging data according to claim 2, characterized in that: In step (3), in the STACGN fusion network, [X 1,1 ...X t,m ] represents two-dimensional sample data with t rows and m columns; [α1...α m ] is the weight vector of the spatial attention layer; α m is the weight of the mth spatial feature; [T 1,1 ...T t,m ] is the feature extracted from the 2DCNN layer; T t,m is the feature value extracted by the mth convolutional layer; [T 1,1 α1...T t,m α m ] represents the output of the spatial attention layer; h t,n1 , H t,n2 The hidden states of the first and second layer GRUs at time step t have n1 features and n2 features respectively; [β1...β t ] is the weight vector of the temporal attention layer; β t is the weight of the feature at the tth time step; [H' 1,1 β1...H' t,n2 β t ] is the output of the temporal attention layer.

4. The method for predicting shale formation cyclic characteristics based on shear wave velocity and density logging data according to claim 3, wherein: In step (4), conventional logging curves include natural gamma curves, compressional wave velocity curves, neutron porosity curves and natural potential curves.

5. The method for predicting shale formation cyclic characteristics based on shear wave velocity and density logging data according to claim 4, characterized in that: In step (4), the STACGN fusion network combines convolutional neural networks and gated recurrent neural units, uses convolutional layers to extract spatial features of logging data, and captures the temporal dependencies of sequence data through GRU layers; During model training, the mean squared error (MSE) was used as the loss function, and the Adam optimizer was used for parameter updates. During training, the Dropout mechanism was introduced, and the learning rate was adjusted. The performance of the model was evaluated through cross-validation, and a low error was achieved on the training set. The loss function is as follows: Where MSE is the mean square error, n is the total number of samples, and Y i is the true value of the i-th sample in the dataset, is the predicted value for this sample.

6. The method for predicting shale formation cyclic characteristics based on shear wave velocity and density logging data according to claim 5, characterized in that: In step (5), R 2 The coefficient of determination is used to measure the degree of fit between the model's predicted value and the actual value. The closer its value is to 1, the stronger the model's ability to explain the data. The expression is as follows: Among them, y i is the true label, is the mean of the true values, is the predicted value of the network, and n is the number of samples.

Images