Volcanic reservoir multi-scale fracture prediction method based on intelligent fusion strategy

Abstract

The present application relates to the field of oil and gas exploration and development, and in particular to a method for predicting multi-scale fractures of volcanic reservoirs based on intelligent fusion strategies. The method comprises the following steps: obtaining five-dimensional seismic data collected at multiple azimuths and multiple incidence angles; standardizing the five-dimensional seismic data to obtain a dataset with multi-scale fracture seismic attributes; inputting the dataset with multi-scale fracture seismic attributes into a pre-trained bidirectional recurrent-convolutional neural network model to obtain a dataset representing multi-scale fracture seismic attributes after fusion of multiple azimuths and multiple incidence angles; standardizing the dataset representing multi-scale fracture seismic attributes after fusion of multiple azimuths and multiple incidence angles; and inputting the standardized dataset representing multi-scale fracture seismic attributes after fusion of multiple azimuths and multiple incidence angles into a pre-trained gated recurrent neural network model to obtain a dataset representing fused multi-scale fracture seismic attributes.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development, specifically to a multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy. Background Technology

[0002] In volcanic rock oil and gas reservoirs, the influencing factors, formation mechanisms, hierarchical combinations, distribution patterns, and control effects on oil and gas of multi-scale fractures differ from those of conventional reservoirs. Therefore, higher requirements are placed on the research of multi-scale fracture identification and characterization.

[0003] Currently, the identification and characterization of multi-scale fractures are trending towards greater refinement and quantification, gradually forming a comprehensive approach to characterizing reservoir multi-scale fractures using multiple methods and parameters. Comprehensive characterization of reservoir multi-scale fractures is the foundation for reservoir multi-scale fracture prediction and modeling, and also an important reference for reservoir engineering sweet spot selection and exploration and development plan formulation. Due to the complexity of the formation mechanisms and influencing factors of multi-scale fractures, it is difficult to accurately depict the three-dimensional distribution of multi-scale fractures using a single method or parameter.

[0004] Furthermore, the accuracy and applicability of identifying fractures using fracture-sensitive attributes are limited, mainly applicable to large-scale main fractures. For medium- and small-scale fractures, the accuracy of fracture prediction is low due to a lack of sufficient resolution and continuity; relying solely on a single attribute for fracture prediction yields low reliability. In addition, fracture interpretation has long been primarily manual, severely impacting work efficiency.

[0005] In the prior art, patent document CN 113138407 A discloses a technical solution for "a multi-scale fracture earthquake prediction method and system for deep shale gas". The steps disclosed in this technical solution are as follows: Step 1: Acquire a three-dimensional seismic time-migration data volume, including drilling data, well logging data, microseismic data, and target layer horizon data; Step 2: Filter the three-dimensional seismic time-migration data volume along the structural direction to obtain a filtered three-dimensional seismic data volume; Step 3: Calculate multiple seismic attributes using the filtered three-dimensional seismic data volume to obtain multiple attribute volumes; Step 4: Perform unsupervised clustering analysis based on the multiple attribute volumes to obtain fracture seismic facies; Step 5: Reconstruct the waveform components of the target layer using the filtered three-dimensional seismic data volume to obtain a reconstructed seismic data volume; Step 6: Enhance the attributes of the reconstructed seismic data volume to obtain attribute-enhanced data; Step 7: Fuse and analyze the fracture seismic facies and attribute-enhanced data to obtain the multi-scale fracture earthquake prediction result for deep shale gas. The disadvantages of this technical solution are: ① The data volume is only three-dimensional seismic data, lacking five-dimensional seismic data acquired from multiple azimuth angles and multiple incident angles; ② Existing technologies still use conventional algorithms in attribute fusion methods, resulting in the loss of effective information; ③ The accumulation of errors in multi-step data processing leads to low prediction accuracy. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy, which is easier to apply in practical engineering and has higher theoretical accuracy. This method can break through the limitations of single methods, thereby identifying fractures at more scales, improving the resolution and continuity of fracture prediction, and progressively improving the accuracy and efficiency of multi-scale fault interpretation.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy, comprising:

[0008] Acquire five-dimensional seismic data from multiple azimuths and incident angles;

[0009] The five-dimensional seismic data is standardized to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults.

[0010] The dataset containing the seismic attributes of large-scale, medium-scale, and small-scale faults is input into a pre-trained bidirectional recurrent-convolutional neural network model to obtain a dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after multi-azimuth and multi-incident angle fusion.

[0011] The dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after the fusion of multi-azimuth and multi-incident angles is standardized.

[0012] The standardized dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after multi-azimuth and multi-incident angle fusion is input into a pre-trained gated recurrent neural network model to obtain the dataset representing the fused seismic attributes of large-scale, medium-scale, and small-scale faults.

[0013] In a preferred embodiment, the step of standardizing the five-dimensional seismic data to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults includes:

[0014] The five-dimensional seismic data is then subjected to guided filtering.

[0015] Based on the filtered five-dimensional seismic data, the coherence, curvature and gradient structure tensors are obtained to characterize the seismic properties of large-scale, medium-scale and small-scale faults, respectively.

[0016] The obtained seismic attribute data of the three types were standardized by z-score, with a mean of 0 and a variance of 1, to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults.

[0017] In a preferred embodiment, the bidirectional recurrent-convolutional neural network model architecture includes: an input layer, four one-dimensional convolutional layers, a bidirectional recurrent neural network layer, and two fully connected layers.

[0018] In a preferred embodiment, the method further includes: a step of training a bidirectional recurrent-convolutional neural network model; the step of training the bidirectional recurrent-convolutional neural network model includes:

[0019] Obtain the first sample dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults;

[0020] The first sample dataset is input into the bidirectional recurrent-convolutional neural network model for training to obtain the pre-trained bidirectional recurrent-convolutional neural network model.

[0021] In a preferred embodiment, the step of obtaining a first sample dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults includes:

[0022] A portion of the dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults is cropped to form the first sample dataset, which is then divided into a first training set and a first validation set.

[0023] In the step of inputting the first sample dataset into the bidirectional recurrent convolutional neural network model for training to obtain the pre-trained bidirectional recurrent convolutional neural network model:

[0024] When training the bidirectional recurrent-convolutional neural network model, the optimizer is set to Nadam, the batch size is 32, the loss function is MSE, and the number of iterations is set to 500 epochs.

[0025] In a preferred embodiment, the step of standardizing the dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after fusing the multi-azimuth and multi-incident angles includes:

[0026] The dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after the fusion of multi-azimuth and multi-incident angles is subjected to z-score standardization, with a mean of 0 and a variance of 1.

[0027] In a preferred embodiment, the gated recurrent neural network model architecture includes: a hidden layer with gated recurrent units and a fully connected layer.

[0028] In a preferred embodiment, the method further includes: a step of training a gated recurrent neural network model; the step of training the gated recurrent neural network model includes:

[0029] Obtain a second sample dataset that represents the seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth and multiple incident angles.

[0030] The second sample dataset is input into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model.

[0031] In a preferred embodiment, the step of obtaining a second sample dataset representing seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth and incident angles includes:

[0032] A portion of the standardized multi-azimuth and multi-incident angle fusion dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults is cropped to serve as a second sample dataset, which is then divided into a second training set and a second validation set.

[0033] In the step of inputting the second sample dataset into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model:

[0034] When training the gated recurrent neural network model, the optimizer is set to Adamax, the training batch size is 64, the loss function is MSE, the number of channels in the GRU is 32, and the number of iterations is 500.

[0035] Beneficial effects of this invention:

[0036] This invention introduces five-dimensional seismic data acquired from multiple azimuth angles and multiple incident angles, which can obtain richer seismic attribute information, thereby improving the ability to identify faults at different scales. The multi-azimuth angle and multiple incident angle seismic attribute fusion method based on the bidirectional recurrent-convolutional neural network (Bi-R&CNN) model can more effectively extract and fuse multi-source information and improve feature representation capabilities.

[0037] The gated recurrent neural network (GRNN) model based on gated recurrent units (GRU) proposed in this invention achieves efficient fusion of multiple attributes, has a simple process, small error accumulation, and improves the accuracy and continuity of fracture prediction. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy, provided in an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the Bi-R&CNN network architecture used for multi-azimuth and multi-incident angle fusion.

[0040] Figure 3 This is a schematic diagram of a GRNN network architecture for multi-attribute fusion.

[0041] Figure 4 A plan view for predicting the degree of multi-scale fracture development in volcanic reservoirs. Detailed Implementation

[0042] The detailed description and technical content of the present invention are explained below with reference to the accompanying drawings. However, the drawings are provided for reference and illustration only and are not intended to limit the present invention.

[0043] like Figure 1 As shown in the figure, the multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy provided by this invention includes:

[0044] Step S1: Acquire five-dimensional seismic data from multiple azimuth angles and multiple incident angles;

[0045] Step S2: Standardize the five-dimensional seismic data to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults.

[0046] Step S3: Input the dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults into a pre-trained bidirectional recurrent-convolutional neural network (Bi-R&CNN) model to obtain a dataset representing earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults after multi-azimuth angle and multi-incident angle fusion.

[0047] Step S4: Standardize the dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after fusing the multi-azimuth and multi-incident angle data.

[0048] Step S5: Input the standardized multi-azimuth and multi-incident angle fusion dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults into a pre-trained gated recurrent neural network (GRNN) model to obtain a fused dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults. This dataset is used to predict the degree of multi-scale fault development in volcanic reservoirs, delineate the boundaries of multi-scale fault development zones in volcanic reservoirs, and statistically analyze the degree of multi-scale fault development in volcanic reservoirs.

[0049] Among them, the five-dimensional seismic data acquired from multiple azimuth angles and multiple incident angles include: (1) seismic data acquired from 0-30° azimuth angle and 0-12° incident angle; (2) seismic data acquired from 0-30° azimuth angle and 12-24° incident angle; (3) seismic data acquired from 0-30° azimuth angle and 24-36° incident angle; (4) seismic data acquired from 30-60° azimuth angle and 0-12° incident angle; (5) 30 Seismic data acquired at -60° azimuth and 12-24° incident angle; (6) Seismic data acquired at 30-60° azimuth and 24-36° incident angle; (7) Seismic data acquired at 60-90° azimuth and 0-12° incident angle; (8) Seismic data acquired at 60-90° azimuth and 12-24° incident angle; (9) Seismic data acquired at 60-90° azimuth and 24-36° incident angle; 10) Seismic data acquired at 90-120° azimuth and 0-12° incident angle; (11) Seismic data acquired at 90-120° azimuth and 12-24° incident angle; (12) Seismic data acquired at 90-120° azimuth and 24-36° incident angle; (13) Seismic data acquired at 120-150° azimuth and 0-12° incident angle; (14) Seismic data acquired at 120-150° azimuth and 12- Seismic data acquired at an incident angle of 24°; (15) Seismic data acquired at azimuth angles of 120-150° and incident angles of 24-36°; (16) Seismic data acquired at azimuth angles of 150-180° and incident angles of 0-12°; (17) Seismic data acquired at azimuth angles of 150-180° and incident angles of 12-24°; (18) Seismic data acquired at azimuth angles of 150-180° and incident angles of 24-36°.

[0050] In some embodiments, step S2, which involves standardizing the five-dimensional seismic data to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults, includes:

[0051] Step S21: Perform guided filtering processing on the five-dimensional seismic data;

[0052] Step S22: Based on the filtered five-dimensional seismic data, the coherence, curvature, and gradient structure tensors are obtained to characterize the seismic properties of large-scale, medium-scale, and small-scale faults, respectively.

[0053] Step S23: The obtained three types of seismic attribute data are subjected to z-score standardization, with a mean of 0 and a variance of 1, to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults.

[0054] Among them, the guided filtering process, the calculation of coherence, curvature and gradient structure tensors, and the z-score normalization can all be performed using existing methods.

[0055] In some embodiments, such as Figure 2 As shown, the architecture of the Bidirectional Recurrent-Convolutional Neural Network (Bi-R&CNN) model includes: one input layer, four one-dimensional convolutional layers, one bidirectional recurrent neural network layer, and two fully connected layers. The last layer serves as the model's output, achieving the final fusion.

[0056] The process begins by using standardized seismic attributes (multi-azimuth and multi-incident angles) as the input vector, i.e., the dataset obtained in step S2. This input vector is then fed into the input layer and propagated through four convolutional layers (using ReLU activation). These convolutional layers extract features from the input data and compress the data size. Each convolutional layer contains a kernel with a number of channels. The kernels in the first three convolutional layers are 2×1 in size, and padding is used to keep the vector size constant. The number of channels in the vector increases from 1 to 64 in the first convolutional layer and remains at 64 throughout the convolutions of the first three layers. In the fourth convolutional layer, padding is removed, resulting in a 1×1 vector with 128 channels.

[0057] The vector form of the input data can be expressed as: ,in It is the length of the input vector. A combination of convolutional kernels can be represented as: The input data undergoes a convolution operation. Obtain the feature mapping:

[0058] (1),

[0059] in, It is an activation function that makes the mapping relationship non-linear. Non-linearity can adjust or cut off the resulting output. Let's define it as ReLU. ReLU has a simple definition in terms of function and gradient:

[0060] (2),

[0061] (3).

[0062] After feature extraction through four convolutional layers, the data size is compressed to 1×1, and the number of channels is expanded to 128. This can be viewed as the features of a long sequence, thus a bidirectional recurrent neural network (RNN), suitable for processing sequence information, is added. RNNs provide an elegant way to learn feature information in a sequence, reflecting the correlation between neighboring data points.

[0063] To make all input information available, two separate RNNs with different directions are used to learn from the long sequence input. Then, the outputs of these two RNNs are arithmetically averaged. During the forward propagation of the BRNN, for a given... The input sequence is passed forward only by the forward RNN (from...). arrive Reverse RNN propagates backward (from...) arrive During the backpropagation process of calculating the descent gradient, the forward RNN updates the weight state backward (from...). arrive The reverse RNN updates the weight state forward (from...). arrive Each direction of the BRNN layer sets 128 hidden states, resulting in 256 hidden states in the proposed model.

[0064] Finally, the feature vectors extracted by the BRNN are passed to the fully connected layers. The first fully connected layer has 1000 hidden units to receive these features. The second fully connected layer has only one hidden unit, corresponding to the seismic attributes fused from multiple azimuth and incident angles. The calculation method of the fully connected layers is as follows:

[0065] (4),

[0066] in, It is a weight matrix. It is the output of the previous layer. It is a weight matrix. It is the sigmoid activation function.

[0067] In some embodiments, the method further includes: step S2a, training a bidirectional recurrent-convolutional neural network model; step S2a, training a bidirectional recurrent-convolutional neural network model, includes:

[0068] Obtain the first sample dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults;

[0069] The first sample dataset is input into the bidirectional recurrent-convolutional neural network model for training to obtain the pre-trained bidirectional recurrent-convolutional neural network model.

[0070] In some embodiments, in the step of obtaining a first sample dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults:

[0071] A portion of the dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults is cropped to form the first sample dataset. The first sample dataset is divided into a first training set and a first validation set (e.g., 20% of the first sample dataset is used as the first validation set).

[0072] In the step of inputting the first sample dataset into the bidirectional recurrent convolutional neural network model for training to obtain the pre-trained bidirectional recurrent convolutional neural network model:

[0073] When training the bidirectional recurrent-convolutional neural network model, the optimizer was set to Nadam, the batch size to 32, the loss function to MSE, and the number of iterations to 500 epochs to provide consistent comparisons.

[0074] Adam is one of the most popular and widely used optimizers today; Nesterov momentum is a key improvement for acceleration and look-ahead; Nadam = Nesterov momentum + Adam. Nadam combines the adaptive learning rate advantage of Adam with the "look-ahead" acceleration advantage of Nesterov momentum, making it an Adam optimizer with Nesterov look-ahead vision.

[0075] In some embodiments, step S4, which involves standardizing the dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after fusing multiple azimuth and multiple incident angles, includes:

[0076] The dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after the fusion of multi-azimuth and multi-incident angles is subjected to z-score standardization, with a mean of 0 and a variance of 1.

[0077] In some embodiments, such as Figure 3 As shown, the gated recurrent neural network (GRNN) model architecture includes: a hidden layer with gated recurrent units (GRUs) and a fully connected layer.

[0078] The model's input data consists of seismic attributes representing large, medium, and small-scale faults after multi-azimuth and multi-incident angle fusion, i.e., a sequence of sample points for each seismic attribute. After standardization, the input data is passed to the hidden layer. In the GRU structure of the hidden layer, two activation functions, sigmoid and tanh, are used. The output of the sigmoid function maps to the interval (0, 1), is monotonically continuous, stable in optimization, and easy to differentiate. The tanh function has a mean of 0 and converges faster than sigmoid. Overall, GRNN replaces each hidden unit in a typical RNN with a GRU, thus achieving the memorization and forgetting of information flow. Each GRU needs to receive the hidden state of the previous unit and input data for updating; therefore, the first GRU needs to utilize... Initialize the unit state. This is the hidden state used to output to the next GRU after the first GRU finishes processing the data. This information flow continues until the last GRU. The prediction for each element is calculated based on the hidden state of the previous element. Finally, the output is data representing the fused seismic attributes of large, medium, and small-scale faults, after multi-azimuth and angular fusion.

[0079] In some embodiments, the method further includes: step S4a, training a gated recurrent neural network model; step S4a, training a gated recurrent neural network model includes:

[0080] Obtain a second sample dataset that represents the seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth and multiple incident angles.

[0081] The second sample dataset is input into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model.

[0082] In some embodiments, the step of obtaining a second sample dataset representing seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth angles and multiple incident angles includes:

[0083] A portion of the standardized multi-azimuth and multi-incident angle fusion dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults is cropped to form a second sample dataset. The second sample dataset is divided into a second training set and a second validation set (e.g., 20% of the second sample dataset is used as the second validation set).

[0084] In the step of inputting the second sample dataset into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model:

[0085] When training the gated recurrent neural network model, the optimizer is set to Adamax, the training batch size is 64, the loss function is MSE, the number of channels of the GRU (which also represents the number of gated units in the recurrent network layer) is 32, and the number of iterations is 500.

[0086] Adamax is a variant of the Adam optimizer.

[0087] Figure 4 This is a plan view for predicting the development degree of multi-scale faults in volcanic reservoirs. Based on five-dimensional seismic data acquired from multiple azimuths and incident angles in a certain region, the method of this invention can be used to obtain a dataset fused from the seismic attributes representing large-scale, medium-scale, and small-scale faults in that region, and thus obtain... Figure 4 The diagram shown is a prediction of the multi-scale fault development in the volcanic reservoirs of this region. As can be seen from the diagram, the fusion results highlight the common characteristics of each attribute, and the effect is good. Medium and small-scale faults develop along large-scale faults, or develop between a group of large-scale faults, which is consistent with geological understanding. This diagram shows that the method of the present invention can improve the ability to identify faults of different scales, and the accuracy and continuity of fault prediction are both high.

[0088] The embodiments described above are merely preferred embodiments for fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A multi-scale fracture prediction method for volcanic reservoirs based on an intelligent fusion strategy, characterized in that, include: Acquire five-dimensional seismic data from multiple azimuths and incident angles; The five-dimensional seismic data is standardized to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults. The dataset containing the seismic attributes of large-scale, medium-scale, and small-scale faults is input into a pre-trained bidirectional recurrent-convolutional neural network model to obtain a dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after multi-azimuth and multi-incident angle fusion. The dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after the fusion of multi-azimuth and multi-incident angles is standardized. The standardized multi-azimuth and multi-incident angle fusion dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults is input into a pre-trained gated recurrent neural network model to obtain the fused dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults. The step of standardizing the five-dimensional seismic data to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults includes: The five-dimensional seismic data is then subjected to guided filtering. Based on the filtered five-dimensional seismic data, the coherence, curvature and gradient structure tensors are obtained to characterize the seismic properties of large-scale, medium-scale and small-scale faults, respectively. The obtained seismic attribute data of the three types were standardized by z-score, with a mean of 0 and a variance of 1, to obtain a dataset with seismic attributes of large-scale faults, medium-scale faults, and small-scale faults.

2. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 1, characterized in that, The bidirectional recurrent-convolutional neural network model architecture includes: one input layer, four one-dimensional convolutional layers, one bidirectional recurrent neural network layer, and two fully connected layers.

3. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 2, characterized in that, Also includes: Steps for training a bidirectional recurrent-convolutional neural network model; The steps for training the bidirectional recurrent-convolutional neural network model include: Obtain the first sample dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults; The first sample dataset is input into the bidirectional recurrent-convolutional neural network model for training to obtain the pre-trained bidirectional recurrent-convolutional neural network model.

4. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 3, characterized in that, In the step of obtaining the first sample dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults: A portion of the dataset with earthquake attributes of large-scale faults, medium-scale faults, and small-scale faults is cropped to form the first sample dataset, which is then divided into a first training set and a first validation set. In the step of inputting the first sample dataset into the bidirectional recurrent convolutional neural network model for training to obtain the pre-trained bidirectional recurrent convolutional neural network model: When training the bidirectional recurrent-convolutional neural network model, the optimizer is set to Nadam, the batch size is 32, the loss function is MSE, and the number of iterations is set to 500 epochs.

5. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 1, characterized in that, The standardization process for the dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after fusing the multi-azimuth and multi-incident angle data includes: The dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after the fusion of multi-azimuth and multi-incident angles is subjected to z-score standardization, with a mean of 0 and a variance of 1.

6. The method for multi-scale fracture prediction of volcanic reservoirs based on intelligent fusion strategy as described in claim 1 or 5, characterized in that, The gated recurrent neural network model architecture includes: a hidden layer with gated recurrent units and a fully connected layer.

7. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 6, characterized in that, Also includes: Steps for training a gated recurrent neural network model; The steps for training the gated recurrent neural network model include: Obtain a second sample dataset that represents the seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth and multiple incident angles. The second sample dataset is input into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model.

8. The method for predicting multi-scale fractures in volcanic reservoirs based on an intelligent fusion strategy as described in claim 7, characterized in that, In the step of obtaining the second sample dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults after fusion of multiple azimuth and incident angles: A portion of the standardized multi-azimuth and multi-incident angle fusion dataset representing the seismic attributes of large-scale, medium-scale, and small-scale faults is cropped to serve as a second sample dataset, which is then divided into a second training set and a second validation set. In the step of inputting the second sample dataset into the gated recurrent neural network model for training to obtain the pre-trained gated recurrent neural network model: When training the gated recurrent neural network model, the optimizer is set to Adamax, the training batch size is 64, the loss function is MSE, the number of channels in the GRU is 32, and the number of iterations is 500.

Images