A cross-block shale gas reservoir parameter prediction method based on MSDC-HiLoTransformer model

Abstract

The present application belongs to the technical field of artificial intelligence and shale gas reservoir evaluation, and particularly relates to a cross-block shale gas reservoir parameter prediction method based on an MSDC-HiLoTransformer model. The present application integrates a multi-scale deep separable convolution module with a DC-HiLoTransformer encoder, and combines three mechanisms of field adversarial training, multi-layer correlation alignment and consistency regularization to construct an MSDC-HiLoTransformer model. The present application uses well logging data and reservoir experimental parameters screened according to correlation as training data to complete model training and perform shale gas reservoir parameter prediction. The present application can solve the problems of weak modeling capability for complex geological features, unstable training under the condition of scarce data in new blocks, and poor model generalization capability caused by distribution deviation of cross-block well logging data in the prior art, and improve the prediction accuracy of shale gas reservoir parameters.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and shale gas reservoir evaluation technology, specifically involving a cross-block shale gas reservoir parameter prediction method based on the MSDC-HiLoTransformer model, aiming to improve the accuracy of intelligent prediction. Background Technology

[0002] Accurate acquisition of shale gas reservoir parameters is crucial for identifying favorable stratigraphic intervals and evaluating geological sweet spots. Traditional methods rely on core experiments, which, while reliable, are costly, time-consuming, and limited in the availability of core samples from new blocks, making them unsuitable for large-scale, rapid evaluation. Existing reservoir parameter prediction methods mainly include rock physics models, empirical formulas, traditional machine learning models, and deep learning models. Rock physics models (such as the ΔlogR method) rely on regional empirical relationships and are easily affected by mineral composition and reservoir heterogeneity, making it difficult to characterize the nonlinear relationship between logging response and reservoir parameters. Empirical formulas are simple to calculate but have poor generalization ability. Traditional machine learning models, such as Random Forest (RF), K-Nearest Neighbors (KNN), and Extreme Gradient Boosting (XGBoost), often exhibit underfitting or noise overfitting when dealing with the complex coupled features of logging data. In recent years, deep learning models have demonstrated significant advantages in reservoir parameter prediction due to their powerful nonlinear modeling capabilities. Convolutional Neural Networks (CNNs) excel at extracting local features; Recurrent Neural Networks (RNNs) and their improved models, such as Long Short-Term Memory Networks (LSTM), Gated Recurrent Units (GRUs), and Bidirectional Long Short-Term Memory Networks (BiLSTM), can capture sequence variation patterns; and the Transformer architecture based on self-attention mechanisms can capture global dependencies. These deep learning models have been widely applied to reservoir parameter prediction, outperforming traditional methods. However, existing deep learning methods still face three challenges in cross-block reservoir parameter prediction:

[0003] (1) Insufficient multi-scale geological feature modeling capability: Existing models are unable to simultaneously take into account the local detailed geological features and global trend geological information contained in well logging data, resulting in low model prediction accuracy.

[0004] (2) The scarcity of samples leads to unstable training: The core samples of the new block are scarce, and the existing model is prone to overfitting when trained on small sample data, resulting in unstable model training and poor generalization ability.

[0005] (3) Cross-block data distribution offset: The distribution of logging data between different blocks is very different. Existing models are difficult to extract domain-invariant features. Direct transfer learning applications will lead to a significant reduction in prediction performance or even negative transfer.

[0006] Therefore, there is an urgent need in this field for a new method that can solve the above problems and achieve high-precision prediction of shale gas reservoir parameters across blocks. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model, comprising the following steps:

[0008] S1. Obtain source and target domain logging data and reservoir experimental parameters, perform correlation screening and preprocessing, and create model training data and model prediction data;

[0009] S2. Integrating the concept of multi-scale and depthwise separable convolution, a multi-scale depthwise separable convolution module is designed.

[0010] S3. Based on the Transformer encoder structure, the DC-HiLo attention mechanism is used to replace the multi-head attention mechanism to form a DC-HiLoTransformer encoder.

[0011] S4 integrates a multi-scale depthwise separable convolutional module with a DC-HiLoTransformer encoder, and combines three mechanisms—domain adversarial training, multi-layer correlation alignment, and consistency regularization—to construct the MSDC-HiLoTransformer model.

[0012] S5. Train the MSDC-HiLoTransformer model using the training data obtained from the preprocessing in step S1.

[0013] S6. Apply the trained MSDC-HiLoTransformer model to predict reservoir parameters in the target block.

[0014] Furthermore, the design scheme for the multi-scale depthwise separable convolutional module in step S2:

[0015] (1) Perform depthwise convolution operation on each channel of the input feature, and use 1×1 pointwise convolution to fuse cross-channel information of the depthwise convolution result to construct a depthwise separable convolution module;

[0016] (2) Depth separable convolution with kernel sizes of 1×1, 3×1 and 5×1 is used respectively to extract geological features at the scales of 1 data point, 3 data points and 5 data points contained in well logging data and reservoir experimental parameters. The three feature maps are spliced ​​and fused together, and then nonlinearly transformed by the ReLU activation function to obtain the output features of the multi-scale depth separable convolution module.

[0017] Furthermore, the DC-HiLoTransformer encoder design scheme in step S3:

[0018] (1) The output of the multi-scale depthwise separable convolutional module after layer normalization is transformed by nonlinear transformation of depthwise separable convolution and GELU activation function to obtain the high-frequency attention mechanism; the output of the multi-scale depthwise separable convolutional module after layer normalization is subjected to average pooling downsampling and then subjected to scaling dot product attention calculation to obtain the low-frequency attention mechanism; the high-frequency attention mechanism and the low-frequency attention mechanism are combined according to learnable weight coefficients. Weighted fusion is performed, followed by linear projection and dropout layer processing to construct the DC-HiLo attention mechanism;

[0019] (2) The DC-HiLo Transformer encoder is constructed by sequentially connecting the layer normalization, DC-HiLo attention mechanism, feedforward network and dropout layer, and introducing residual connection structure after DC-HiLo attention mechanism and feedforward network respectively.

[0020] Furthermore, the MSDC-HiLoTransformer model is designed as follows in step S4:

[0021] (1) Front-end feature extraction: The linear projection branch and the multi-scale depthwise separable convolutional module are weighted by learnable weight coefficients. Weighted fusion is performed, and positional encoding is introduced to obtain the front-end embedded representation;

[0022] (2) Feature encoding: The front-end embedded representation is used to input the output features. Feature extraction is performed in a layer-stacked DC-HiLoTransformer encoder;

[0023] (3) Domain adaptive alignment: Apply two random data augmentations to the input features of the target domain and apply consistency constraints to the same feature under different augmentation conditions; calculate the feature covariance matrix of the source domain and the target domain in multiple intermediate layers, and align the second-order statistics by minimizing the covariance difference; set a domain classifier at the encoder output and connect it through a gradient inversion layer so that the feature extractor learns the domain-invariant feature representation.

[0024] (4) Predicted output: Based on the output features of the encoder, a fixed-length feature vector is obtained through global average pooling, and then input into the fully connected layer to obtain the model output.

[0025] Furthermore, the model described in S5 is trained through the following steps:

[0026] (1) Input the source domain and target domain logging data and reservoir experimental parameters into the model, extract deep features through forward propagation, and update the model parameters iteratively through backpropagation based on the source domain reservoir experimental parameters to complete the model pre-training;

[0027] (2) Freeze the parameters of the pre-trained model backbone network, and use only a small number of reservoir experimental parameters in the target domain to backpropagate and update the parameters of the fully connected layer to complete the model fine-tuning and obtain the trained MSDC-HiLoTransformer model.

[0028] The present invention has the following beneficial effects:

[0029] (1) This invention constructs a multi-scale depth-separable convolution module, splices and fuses depth-separable convolutions of different scales (such as 1×1, 3×1, 5×1), and introduces a learnable weighted fusion mechanism and linear projection branch to achieve adaptive extraction of local geological features in well logging curves. On this basis, combined with the improved HiLoTransformer collaborative modeling of global trends and local details, it effectively solves the problem of weak modeling ability of existing technologies for complex geological features and significantly improves the prediction accuracy of reservoir parameters.

[0030] (2) This invention embeds a DC-HiLo attention mechanism into the Transformer encoder, which integrates depthwise separable convolution with high and low frequency attention, enhancing the model's ability to represent different frequency components in well logging data. At the same time, a two-stage transfer learning strategy is designed, integrating three mechanisms: domain adversarial training, multi-layer correlation alignment, and consistency regularization, to gradually reduce the distribution difference between the source and target domains. This effectively alleviates the problem of poor model generalization ability caused by cross-block well logging response differences and data distribution shifts, and can still maintain stable prediction performance under the condition of scarce data in new blocks.

[0031] (3) This invention can effectively solve the small sample learning dilemma caused by the scarcity of core samples in new blocks by using the efficient parameter design of multi-scale depth separable convolution modules and the domain adaptive transfer learning mechanism. It can achieve effective transfer and reuse of source domain knowledge under the condition of a small number of target domain samples, and provide reliable technical support for shale gas reservoir evaluation and exploration and development. Attached Figure Description

[0032] Figure 1 This is a flowchart of the cross-block shale gas reservoir parameter prediction method based on the MSDC-HiLoTransformer model described in this invention;

[0033] Figure 2 This is a cross-plot of well logging data and reservoir experimental parameters according to an embodiment of the present invention;

[0034] Figure 3 This invention relates to a multi-scale depth-separable convolution module design scheme;

[0035] Figure 4 This is the DC-HiLoTransformer encoder design scheme of the present invention;

[0036] Figure 5 This is a structural diagram of the MSDC-HiLoTransformer model of the present invention;

[0037] Figure 6 This is a curve comparing the experimental test value and the predicted value of organic carbon content in well C1 of this embodiment of the invention.

[0038] Figure 7 This is a curve comparing the experimental test value and the predicted value of organic carbon content in well C2 of this invention.

[0039] Figure 8 This is a curve comparing the experimental test value and the predicted value of organic carbon content in well C3 of this invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0041] A method for predicting parameters of cross-block shale gas reservoirs based on the MSDC-HiLoTransformer model includes the following steps, as shown in the appendix. Figure 1 As shown:

[0042] S1: Obtain source and target domain logging data and reservoir experimental parameters, perform correlation screening and preprocessing, and create model training and model prediction data;

[0043] Well logging data includes natural gamma (GR), uranium-free gamma (KTH), uranium (U), acoustic (AC), and density (DEN); reservoir experimental parameters include total organic carbon (TOC), porosity (POR), and gas content (GAST). Pearson correlation coefficients were used. Correlation analysis was performed on well logging data and reservoir experimental parameters to select well logging data with a high correlation to the predicted reservoir parameters, as shown in the attached figure. Figure 2 As shown, the calculation formula is as follows:

[0044]

[0045] In the formula, Indicates the first One well logging data; Indicates the first Reservoir experimental parameters corresponding to each logging data point; This represents the average value of the well logging data. The mean values ​​of the reservoir experimental parameters; This represents the number of samples.

[0046] The preprocessing steps include:

[0047] (1) Calculate the interquartile range based on quantiles, and select outliers in well logging data and reservoir experimental parameters:

[0048]

[0049]

[0050] In the formula, and They are the 25th and 75th percentiles, respectively. and These are the cutoff intervals. and The cutting factor, The data was then cropped to... interval;

[0051] (2) Denoising was performed on the trimmed logging data and reservoir experimental parameters using smoothing filtering;

[0052] (3) The denoised logging data and reservoir experimental parameters were standardized using the standard score Z-Score:

[0053]

[0054] In the formula, These are the original values ​​before standardization. The mean, Standard deviation These are the standardized values.

[0055] S2, integrating multi-scale concepts with depthwise separable convolution, designs a multi-scale depthwise separable convolution module using the following scheme, as shown in the appendix. Figure 3 As shown:

[0056] (1) Perform depthwise convolution on each channel of the input feature, and use 1×1 pointwise convolution to fuse cross-channel information of the depthwise convolution result to construct a depthwise separable convolution module. The calculation formulas for depthwise convolution, pointwise convolution, and depthwise separable convolution are as follows:

[0057]

[0058]

[0059]

[0060] In the formula, Calculate dimensions from model input Depth-wise convolution operations; It is a 1×1 pointwise convolution operation performed on the model input; Calculate dimensions from model input Depth-separable convolution operations; The kernel size; These are the original input features; This is a convolution operation; and These are the kernel weights for depthwise convolution and pointwise convolution, respectively. and These are the bias terms for depthwise convolution and pointwise convolution, respectively;

[0061] (2) Depth-separable convolution with kernel sizes of 1×1, 3×1, and 5×1 was used to extract geological features at scales of 1, 3, and 5 data points from well logging data and reservoir experimental parameters. The three feature maps were then stitched together and fused, and the output features of the multi-scale depth-separable convolution module were obtained by nonlinear transformation using the ReLU activation function. It can be calculated using the following formula:

[0062]

[0063] In the formula, , , These represent calculating the dimensions of the model input. Separable convolution operations with depths of 3×1 and 5×1; For splicing and merging operations; It is a non-linear activation function.

[0064] S3, based on the Transformer encoder structure, utilizes the DC-HiLo attention mechanism instead of the multi-head attention mechanism to form a DC-HiLoTransformer encoder. The DC-HiLoTransformer encoder is constructed through the following steps, as shown in the attached figure. Figure 4 As shown:

[0065] (1) The output of the multi-scale depthwise separable convolutional module after layer normalization is transformed by nonlinear transformation of depthwise separable convolution and GELU activation function to obtain the high-frequency attention mechanism; the output of the multi-scale depthwise separable convolutional module after layer normalization is subjected to average pooling downsampling and then subjected to scaling dot product attention calculation to obtain the low-frequency attention mechanism; the high-frequency attention mechanism and the low-frequency attention mechanism are combined according to learnable weight coefficients. Weighted fusion is performed, followed by linear projection and a dropout layer to construct a DC-HiLo attention mechanism. The output features of the DC-HiLo attention mechanism are shown below. It can be calculated using the following formula:

[0066]

[0067]

[0068]

[0069] In the formula, These are the input features for high-frequency attention mechanisms. This is a depthwise convolution operation; For pointwise convolution operation, It is a non-linear activation function; For query vector; This represents the transpose operation of the key vector; It is a value vector; This is the scaling factor; The function normalizes the scaled similarity scores into an attention weight distribution; The weighted fusion coefficient; The projection matrix; For discarding layers; and These are the output characteristics of high-frequency attention mechanisms and low-frequency attention mechanisms, respectively.

[0070] (2) The DC-HiLo Transformer encoder is constructed by sequentially connecting the layer normalization, DC-HiLo attention mechanism, feedforward network and dropout layer, and introducing residual connection structure after DC-HiLo attention mechanism and feedforward network respectively.

[0071] S4 integrates a multi-scale depthwise separable convolutional module with a DC-HiLoTransformer encoder, and combines three mechanisms—domain adversarial training, multi-layer relevance alignment, and consistency regularization—to construct the MSDC-HiLoTransformer model, as shown in the attached figure. Figure 5 As shown:

[0072] (1) Front-end feature extraction: The linear projection branch and the multi-scale depthwise separable convolutional module are weighted by learnable weight coefficients. Weighted fusion is performed, and positional encoding is introduced to obtain the front-end embedding representation, and the output features are then generated. Calculated using the following formula:

[0073]

[0074]

[0075] In the formula, It is a linear projection matrix; These are the original input features; It is the bias vector; The weighted fusion coefficient; The output features of the linear projection branch; The output features of a multi-scale depthwise separable convolutional module;

[0076] (2) Feature encoding: The front-end embedded representation is used to input the output features. Feature extraction is performed in a layer-stacked DC-HiLoTransformer encoder. Output features of layer encoder It can be calculated using the following formula:

[0077]

[0078] In the formula, Indicates the first The output features of the layer encoder, where The output features of the front-end feature extraction structure; Indicates the first The mapping function of the layer encoder; For encoder layer index; Indicates the number of layers in the encoder;

[0079] (3) Domain Adaptive Alignment: Apply two random data augmentations to the input features of the target domain, and apply consistency constraints to the same feature under different augmentation conditions; calculate the feature covariance matrix of the source domain and the target domain in multiple intermediate layers, and align the second-order statistics by minimizing the covariance difference; set a domain classifier at the encoder output and connect it through a gradient inversion layer so that the feature extractor learns the domain-invariant feature representation. The loss function of the consistency constraint... Calculated using the following formula:

[0080]

[0081]

[0082]

[0083] In the formula, and Represent two random data augmentation operators, for the target domain... Input data Apply two different random perturbations; Index the target domain samples; Indicated by A nonlinear mapping function constructed for the model parameters; The number of samples in the target domain; and These are the predicted outputs of the target domain input data under two random augmentations, respectively.

[0084] The covariance matrix of the source domain Covariance matrix with the target domain Calculate using the following formulas respectively:

[0085]

[0086]

[0087] In the formula, and These are the feature matrices of the source domain and the target domain, respectively; and These are the number of samples in the source domain and the target domain, respectively. This represents the transpose operation of a vector consisting entirely of 1s.

[0088] Based on the aforementioned covariance matrix, multi-layer covariance alignment loss Calculated using the following formula:

[0089]

[0090]

[0091] In the formula, and These are the covariance matrices of the source domain and the target domain, respectively. For feature dimensions; It is the Fibonacci norm; The total number of layers participating in covariance alignment; Indicates the first Covariance alignment loss corresponding to the layer.

[0092] The domain classifier uses the cross-entropy loss function, and the domain classification loss is calculated using the following formula. :

[0093]

[0094]

[0095]

[0096] In the formula, and These are the number of samples in the source domain and the target domain, respectively. Represents the source domain. One input data; Indicates the target domain. One input data; For domain tags, Represents the source domain. Indicates the target domain; For the domain classifier to the samples The predicted probability is obtained by normalizing the output of the domain classifier using softmax. The domain loss is the loss of the source domain; The domain classification loss is used for the target domain.

[0097] Training loss of the MSDC-HiLoTransformer model It can be calculated using the following formula:

[0098]

[0099]

[0100] In the formula, This represents the number of samples in the source domain. and The source domain is respectively Experimental test values ​​and model prediction values ​​for each sample; Indicates the source domain task regression loss; Representation domain classification loss; This represents the multi-level covariance alignment loss; This represents the loss due to consistency constraints; Weight coefficients for domain classification loss, The weighting coefficients for multi-level covariance alignment loss, The weighting coefficients for the consistency constraint loss.

[0101] (4) Predicted output: based on the output features of the encoder A fixed-length feature vector is obtained through global average pooling, which is then input into a fully connected layer to obtain the model output, which is the feature value. The calculation formula is as follows:

[0102]

[0103] In the formula, and These are the weights and biases of the fully connected layer, respectively. The output characteristics of the DC-HiLoTransformer encoder; This indicates global average pooling.

[0104] S5, using the training data obtained from the preprocessing in step S1, train the MSDC-HiLoTransformer model. The model training is performed through the following steps, as shown in the attached diagram. Figure 1 As shown:

[0105] (1) Input the source domain and target domain logging data and reservoir experimental parameters into the model, extract deep features through forward propagation, and update the model parameters iteratively through backpropagation based on the source domain reservoir experimental parameters to complete the model pre-training;

[0106] (2) Freeze the parameters of the pre-trained model backbone network, and use only a small number of reservoir experimental parameters in the target domain to backpropagate and update the parameters of the fully connected layer to complete the model fine-tuning and obtain the trained MSDC-HiLoTransformer model.

[0107] S6. The trained MSDC-HiLoTransformer model is applied to the reservoir parameter pre-processing of the target block.

[0108] Taking the Wufeng Formation-Longmaxi Formation shale gas reservoir, a key development stratum from the Jiaoshiba Block of the Fuling Shale Gas Field in eastern Sichuan Basin to the Luzhou Shale Gas Field in Sichuan Basin, as an example, the technical solution of this invention is applied to predict the core parameter, total organic carbon (TOC).

[0109] The Jiaoshiba block of the Fuling shale gas field and the Luzhou shale gas field are highly developed. The target layer, the Wufeng Formation-Longmaxi Formation, has abundant core experimental data and well logging data, which can provide reliable prior knowledge for cross-block transfer learning of the MSDC-HiLoTransformer model. Four wells (A1, A2, A3, A4) in the Jiaoshiba block were selected as the source domain, and three wells (C1, C2, C3) in the Luzhou block were selected as the target domain for training and prediction of the MSDC-HiLoTransformer model of this invention.

[0110] MSDC-HiLoTransformer model training parameter configuration: Before data preprocessing, the source domain dataset is divided into training and validation sets in an 8:2 ratio; 30% of each well in the target domain is used as the fine-tuning set, and the remaining 70% is used as the test set. During the pre-training phase, domain-adaptive training is performed using labeled data from the source domain and unlabeled data from the target domain. An early stopping mechanism is introduced on the 20% validation set in the source domain; training terminates when the validation set loss does not decrease for 50 consecutive epochs to avoid overfitting. During the fine-tuning phase, optimization continues on the 30% fine-tuning set of the target wells, and the model weights with the minimum training loss are saved. The MSDC-HiLoTransformer model training parameter configuration is shown in Table 1.

[0111] Table 1 MSDC-HiLoTransformer Model Training Parameter Configuration

[0112]

[0113] The mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) are used to evaluate the predictive performance of the model on the target task. The calculation formulas for each evaluation index are as follows:

[0114]

[0115]

[0116]

[0117]

[0118] In the formula, The number of samples in the test set for the target domain; and They represent the target domain test set number respectively. Experimental test values ​​and model prediction values ​​for a sample.

[0119] Tables 2, 3, and 4 show the performance comparison of the MSDC-HiLoTransformer model with traditional machine learning models (RF, KNN, XGboost), deep learning models (CNN, GRU, Deep Transformer), empirical formulas, and domain adaptive transfer methods on the test set of three wells (C1, C2, and C3) in the target domain. The results show that the MSDC-HiLoTransformer model achieved the lowest prediction error in all three wells. Table 2 shows that the MSDC-HiLoTransformer model achieved an RMSE of 0.2091 in well C1, a reduction of 31.0%-66.4% compared to deep learning models and over 50% compared to classic machine learning models, while also achieving the best performance in all error metrics (MAE, MSE, MAPE). Table 3 shows that the MAE in well C2 was 0.1864, a decrease of approximately 27.8% compared to the second-best performing Deep Transformer. Table 4 shows that the error metrics in well C3 were also the lowest among all compared methods. Overall, the MSDC-HiLoTransformer model consistently achieves the minimum error and delivers the best prediction performance in the organic carbon content prediction task.

[0120] Table 2 Comparison of Organic Carbon Content Prediction Performance in Well C1 of Target Area

[0121]

[0122] Table 3 Comparison of Predictive Performance of Organic Carbon Content in Well C2 in Target Area

[0123]

[0124] Table 4 Comparison of Predictive Performance of Organic Carbon Content in Well C3 in Target Area

[0125]

[0126] Figure 6 , Figure 7 and Figure 8 The experimental and predicted values ​​of the MSDC-HiLoTransformer model, traditional machine learning models, deep learning models, empirical formulas, and domain-free adaptive transfer methods are compared on a test set of three wells (C1, C2, and C3) in the target domain. Overall, the prediction curves of traditional machine learning methods exhibit a tendency to under-predict high values ​​and over-predict low values ​​in the reservoir parameter abrupt change range. Deep learning models outperform traditional models in overall trend fitting, but still suffer from enhanced oscillations and response lag in local segments, especially in the high TOC segment, where the predicted curves still deviate from the actual curves. In contrast, the MSDC-HiLoTransformer model demonstrates the best prediction performance in all three wells.

[0127] In summary, the MSDC-HiLoTransformer model significantly outperforms the comparative models in terms of prediction accuracy and generalization stability, and can achieve reliable prediction of shale gas reservoir parameters across blocks under limited sample conditions.

[0128] Those skilled in the art should understand that the above embodiments are merely illustrative and are not intended to imply that the scope of the invention is limited to these examples. Within the framework of this invention, technical features of the above embodiments or different embodiments can be combined, steps can be implemented in any order, and many other variations of the different aspects of the invention as described above exist, which are not provided in detail for the sake of brevity. Any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting parameters of cross-block shale gas reservoirs based on the MSDC-HiLoTransformer model, characterized in that, Includes the following steps: S1. Obtain source and target domain logging data and reservoir experimental parameters, perform correlation screening and preprocessing, and create model training data and model prediction data; S2. Integrating the concept of multi-scale and depthwise separable convolution, a multi-scale depthwise separable convolution module is designed. S3. Based on the Transformer encoder structure, the DC-HiLo attention mechanism is used to replace the multi-head attention mechanism to form a DC-HiLoTransformer encoder. S4 integrates a multi-scale depthwise separable convolutional module with a DC-HiLoTransformer encoder, and combines three mechanisms—domain adversarial training, multi-layer correlation alignment, and consistency regularization—to construct the MSDC-HiLoTransformer model. S5. Train the MSDC-HiLoTransformer model using the training data obtained from the preprocessing in step S1. S6. Apply the trained MSDC-HiLoTransformer model to predict reservoir parameters in the target block.

2. The method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model according to claim 1, characterized in that, The correlation screening and preprocessing method described in step S1 is as follows: using the Pearson correlation coefficient. Correlation analysis was performed on well logging data and reservoir experimental parameters to screen out well logging data with a high correlation to the predicted reservoir parameters; interquartile ranges were calculated based on quantiles to trim outliers in the screened well logging data and reservoir experimental parameters; smoothing filtering was used to denoise the trimmed well logging data and reservoir experimental parameters; and Z-score was used to standardize the denoised well logging data and reservoir experimental parameters.

3. The method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model according to claim 1, characterized in that, The specific steps for constructing the multi-scale depth-separable convolution module in step S2 are as follows: (1) Perform depth convolution operation on each channel of the input feature, and use 1×1 pointwise convolution to perform cross-channel information fusion on the depth convolution result to construct the depth-separable convolution module. The calculation formulas for depth convolution, pointwise convolution, and depth-separable convolution are as follows: ; ; In the formula, Calculate dimensions from model input Depth-wise convolution operations; It is a 1×1 pointwise convolution operation performed on the model input; Calculate dimensions from model input Depth-separable convolution operations; The kernel size; These are the original input features; This is a convolution operation; and These are the kernel weights for depthwise convolution and pointwise convolution, respectively. and These are the bias terms for depthwise convolution and pointwise convolution, respectively; (2) Depth-separable convolution with kernel sizes of 1×1, 3×1, and 5×1 was used to extract geological features at scales of 1, 3, and 5 data points from well logging data and reservoir experimental parameters. The three feature maps were then stitched together and fused, and the output features of the multi-scale depth-separable convolution module were obtained by nonlinear transformation using the ReLU activation function. Calculated by the following formula: In the formula, , , These represent depthwise separable convolution operations with dimensions of 1×1, 3×1, and 5×1 respectively on the model input; For splicing and merging operations; It is a non-linear activation function.

4. The method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model according to claim 1, characterized in that, The specific steps for constructing the DC-HiLoTransformer encoder in step S3 are as follows: (1) The output of the multi-scale depth-separable convolutional module after layer normalization is subjected to nonlinear transformation by depth-separable convolution and GELU activation function to obtain the high-frequency attention mechanism; the output of the multi-scale depth-separable convolutional module after layer normalization is subjected to average pooling downsampling and then subjected to scaling dot product attention calculation to obtain the low-frequency attention mechanism; the high-frequency attention mechanism and the low-frequency attention mechanism are combined according to learnable weight coefficients. Weighted fusion is performed, followed by linear projection and dropout layer processing to construct a DC-HiLo attention mechanism, whose output features... Calculated using the following formula: ; ; In the formula, These are the input features for high-frequency attention mechanisms. This is a depthwise convolution operation; For pointwise convolution operation, It is a non-linear activation function; For query vector; This represents the transpose operation of the key vector; It is a value vector; This is the scaling factor; The function normalizes the scaled similarity scores into an attention weight distribution; The weighted fusion coefficient; The projection matrix; For discarding layers; and These are the output characteristics of high-frequency attention mechanisms and low-frequency attention mechanisms, respectively. (2) The DC-HiLo Transformer encoder is constructed by sequentially connecting the layer normalization, DC-HiLo attention mechanism, feedforward network and dropout layer, and introducing residual connection structure after DC-HiLo attention mechanism and feedforward network respectively.

5. The method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model according to claim 1, characterized in that, The MSDC-HiLoTransformer model design described in step S4 is as follows: (1) Front-end feature extraction: The linear projection branch and the multi-scale depthwise separable convolutional module are weighted by learnable weight coefficients. Weighted fusion is performed, and positional encoding is introduced to obtain the front-end embedding representation, and the output features are then generated. Calculated using the following formula: ; In the formula, It is a linear projection matrix; These are the original input features; It is the bias vector; The weighted fusion coefficient; The output features of the linear projection branch; The output features of a multi-scale depthwise separable convolutional module; (2) Feature encoding: The front-end embedded representation is used to input the output features. Feature extraction is performed in a layer-stacked DC-HiLoTransformer encoder. Output features of layer encoder Calculated by the following formula: In the formula, Indicates the first The output features of the layer encoder, where The output features of the front-end feature extraction structure; Indicates the first The mapping function of the layer encoder; For encoder layer index; Indicates the number of layers in the encoder; (3) Domain Adaptive Alignment: Apply two random data augmentations to the input features of the target domain, and apply consistency constraints to the same feature under different augmentation conditions; calculate the feature covariance matrix of the source domain and the target domain in multiple intermediate layers, and align the second-order statistics by minimizing the covariance difference; set a domain classifier at the encoder output and connect it through a gradient inversion layer so that the feature extractor learns the domain-invariant feature representation, and the covariance matrix of the source domain is... Covariance matrix with the target domain Calculate using the following formulas respectively: ; In the formula, and These are the feature matrices of the source domain and the target domain, respectively; and These represent the number of samples in the source domain and the target domain, respectively. For feature dimensions; This represents the transpose operation of a vector consisting entirely of 1s; (4) Predicted output: based on the output features of the encoder A fixed-length feature vector is obtained through global average pooling, which is then input into a fully connected layer to obtain the model output, which is the feature value. Calculated by the following formula: In the formula, and These are the weights and biases of the fully connected layer, respectively. Indicates global average pooling; This refers to the output characteristics of the DC-HiLoTransformer encoder.

6. The method for predicting cross-block shale gas reservoir parameters based on the MSDC-HiLoTransformer model according to claim 1, characterized in that, Step S5 is performed as follows: input the source domain and target domain logging data and reservoir experimental parameters into the model, extract deep features through forward propagation, and iteratively update the model parameters through backpropagation based on the source domain reservoir experimental parameters to complete the model pre-training; freeze the pre-trained model backbone network parameters, and use only a small number of reservoir experimental parameters from the target domain to backpropagate and update the fully connected layer parameters to complete the model fine-tuning, thus obtaining the trained MSDC-HiLoTransformer model.

Images