A method for constructing low-frequency models based on spatiotemporally gated cyclic unit fusion networks

Abstract

This invention discloses a method for constructing low-frequency models based on a spatiotemporally gated recurrent unit fusion network, relating to the field of petroleum geophysical exploration. This invention proposes a low-frequency model prediction neural network framework for small sample data. This framework uses GRU as the basic unit to fully consider the long-term dependencies of the low-frequency model's time series, ensuring the integrity of long-wavelength information in the constructed low-frequency model. To improve the horizontal continuity of the low-frequency model, a spatiotemporal feature fusion module is designed to transform multichannel data into potential vectors containing spatial information. Simultaneously, envelope information and the initial model serve as prior information to guide the neural network, helping the network generate stable low-frequency model predictions and reduce prediction nonlinearity, further enhancing the network's generalization ability. It also generates spatial perturbations on the original data to produce label data with richer vertical combinations and more diverse types.

Description

Technical Field

[0001] This invention belongs to the field of petroleum geophysical exploration, specifically a method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network. Background Technology

[0002] Traditional seismic inversion methods aim to construct an objective functional that minimizes the error between the synthesized seismic record and the actual seismic record using forward modeling algorithms. The initial model is continuously updated through optimization algorithms to obtain the optimal solution. However, the actual seismic record lacks low-frequency components. Based on the fact that traditional seismic inversion is heavily dependent on the given initial model, this paper addresses the problems in the construction of conventional initial models by utilizing the superior mapping ability and flexibility of neural networks to construct reasonable low-frequency models, laying the foundation for high-precision seismic inversion. However, in practical applications, neural networks are unstable and have poor generalization ability, making it difficult to promote and apply them on a large scale in production. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and address the problems of complex structures and sparse logging in actual work areas. This invention provides a method for constructing low-frequency models based on spatiotemporally gated cyclic unit fusion networks, which realizes the mining of information with spatial lateral continuity and long-term temporal dependence, and improves the generalization ability of neural networks through multi-information fusion strategies and data augmentation methods.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network, comprising the following steps:

[0005] S1: Conduct wave impedance data augmentation. Based on the original dataset, introduce an elastic transformation algorithm for data augmentation. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through well logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the longitudinal spatial disturbance.

[0006] S2: Establish a low-frequency model sample library. Seismic records are characterized by the convolution of seismic wavelet and stratigraphic reflection coefficient.

[0007] S3: Neural network construction and parameter training. Construct a neural network model that includes a temporal feature extraction module and a spatiotemporal feature fusion module, and construct a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model.

[0008] S4: Neural network test. Calculate the envelope of the seismic data and use the seismic data, initial model, and seismic envelope as input to form a three-channel data set from the well-side i-channel. Similarly, standardize the data to be predicted and use the processed data to conduct neural network tests.

[0009] Preferably, the elastic transformation process consists of three steps: first, a new spatial coordinate is obtained based on the original data coordinate space through affine transformation; then, the new coordinate space is smoothed using a Gaussian function to obtain a random displacement field; finally, the original data is mapped to the new coordinate space to obtain the perturbed data. The elastic transformation process is represented in the following form:

[0010]

[0011] Where D represents the generated random displacement field, T represents the time series, α controls the degree of perturbation of the random displacement field, and δ is the size of the Gaussian filter kernel, which controls the smoothness of the random displacement field. The larger α is and the smaller δ is, the more severe the spatial perturbation. Conversely, the more severe the spatial perturbation, the smoother the spatial perturbation. The elastic transformation algorithm changes the distribution and thickness of reservoir parameters in the original sample, increasing the diversity of the sample.

[0012] Preferably, the seismic record is represented by the convolution of the seismic wavelet and the stratigraphic reflection coefficient, and the relationship between wave impedance and reflection coefficient is as follows:

[0013]

[0014] Among them, t n This represents the nth time sampling point, Δt represents the time increment, R is the reflection coefficient, and Z is the wave impedance;

[0015] The convolution model process is represented as follows:

[0016] S=R*W+N (3)

[0017] Where S represents seismic data, R represents the reflection coefficient, W represents the Ricker wavelet, and N represents noise. The expression for the Ricker wavelet W is as follows:

[0018]

[0019] Where f m The dominant frequency of the seismic wavelet is denoted by t, which represents time. To further expand the diversity of the sample, the dominant frequency of the seismic data is determined by time-frequency analysis. Five dominant frequency information are obtained by adding or subtracting from the dominant frequency of the actual seismic data at 5 Hz intervals.

[0020] Preferably, five dominant frequency information points are obtained by adding or subtracting from the dominant frequency of the actual seismic data at 5Hz intervals, and the seismic record is then subjected to Hilbert transform using the following formula:

[0021]

[0022] Where y(t) represents the earthquake record, y H(t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

[0023] Preferably, a high-pass filter is used to filter out low-frequency components of the synthesized data according to the original data frequency band to simulate the lack of low-frequency components in actual seismic data. The envelope of the band-limited seismic data still carries long-wavelength information of the subsurface medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by calculating the amplitude of the analytic signal. The process is written as follows:

[0024]

[0025] Where e represents the envelope of the seismic signal, the final sample library contains seismic data, seismic envelope, initial model and low-frequency model. Seismic data, seismic envelope and initial model are used as input, and low-frequency model is used as training label. The input data is taken from 5 well-side traces to form 101 pairs of two-dimensional data to enhance the spatial representation ability of the prediction results. The label data is one-dimensional data. Finally, the data is standardized to reduce the difficulty of neural network training. The low-frequency model is standardized using the mean and variance of the initial model.

[0026] Preferably, a single gated recurrent unit (GRU) improves a simple recurrent neural network by coupling a reset gate and an update gate, as shown below:

[0027]

[0028] Among them, u t r t X t , and h t Let represent the updated gate vector, reset gate vector, input sample, candidate hidden state, and hidden state at time t, respectively. w and b represent the weight parameters and bias terms, respectively. Sigmoid and tanh represent different activation functions. * represents the dot product operator.

[0029] Preferably, the bidirectional gated recurrent network (Bi-GRU) unit builds a backpropagation layer on the basis of GRU, taking into account both forward and reverse time dependencies. The temporal feature extraction module consists of multiple parallel 3-in-1 Bi-GRUs. The number of parallel blocks depends on the value of i in step S4. If it is set to 2, there are 5 parallel blocks. The spatiotemporal feature fusion module consists of a Bi-GRU and a fully connected layer.

[0030] Preferably, the parameter training steps of the neural network include: initializing the network weight parameters w and bias term b; inputting the two-dimensional seismic data, initial model, and seismic envelope from the sample library into the neural network; outputting a one-dimensional predicted low-frequency model; and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows:

[0031]

[0032] Where N is the number of sampling points per channel, f is the network model, and Z is the number of sampling points per channel. L This is a practical low-frequency model, where S represents the seismic record and Z represents the actual low-frequency model. init For the initial model, e is the earthquake envelope, W and b represent the network weights and biases, respectively. The first term of the loss function is the low-frequency wave impedance model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

[0033] Preferably, the derivatives of the input data at each step of the neural network are recorded, and the partial derivatives of the loss function with respect to the weight parameters and bias terms are obtained using the chain rule. Then, the Adam optimizer based on variable step size gradient descent is used to calculate the update amount of the weight parameters and bias terms, with the learning rate set to 0.05. The weight parameters and bias terms are updated, and this process is repeated until the error of the loss function is minimized. Finally, the network weight parameters and bias terms are saved.

[0034] The low-frequency model construction method based on spatiotemporally gated recurrent unit fusion network disclosed in this invention also includes a neural network testing method, comprising the following steps:

[0035] (1) When constructing a low-frequency model for test data, first calculate the envelope of the seismic data and take the seismic data, the initial model and the seismic envelope as well-side i-channel data to form three channels as input. Similarly, the data to be predicted is standardized.

[0036] (2) Load the trained network weights and biases to construct a low-frequency model on the test data. Extract the variance σ and mean μ of the initial model to perform inverse standardization on the predicted low-frequency model. The inverse standardization is represented as follows:

[0037] x=σX+μ (9)

[0038] Where X represents standardized data and x represents destandardized data.

[0039] Beneficial effects: The low-frequency model construction method based on spatiotemporal gated cyclic unit fusion network disclosed in this invention is a novel joint inversion structure coupling idea compared with existing known technologies. It uses singular values ​​to characterize the spatial distribution characteristics of multiple physical attributes of the model, constructs an objective function with low-rank structure constraints, realizes joint inversion of multiple geophysical fields, and improves the correlation of multi-physical attribute models. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0041] In the attached diagram:

[0042] Figure 1 This is a flowchart of the low-frequency model construction method based on data augmentation of the present invention;

[0043] Figure 2 This is a schematic diagram of the spatiotemporal gated cyclic fusion network of the present invention;

[0044] Figure 3 This is the original data of Marmousi2 of this invention;

[0045] Figure 4 This invention relates to the dataset augmentation method based on elastic transformation, which utilizes different combinations of α and δ perturbations to achieve various perturbation effects.

[0046] Figure 5 These are the prediction results and absolute errors of different network models in this invention. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following text is only used to describe an implementation method of a low-frequency model construction method based on a spatiotemporally gated cyclic unit fusion network of the present invention, and does not strictly limit the scope of protection specifically claimed by the present invention.

[0048] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they can be implemented by those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope claimed by this utility model.

[0049] Example 1:

[0050] like Figure 1 , Figure 2 and Figure 4As shown, this invention discloses a method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network, comprising the following steps:

[0051] S1: Conduct wave impedance data augmentation. Based on the original dataset, introduce an elastic transformation algorithm for data augmentation. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through well logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the longitudinal spatial disturbance.

[0052] S2: Establish a low-frequency model sample library. Seismic records are characterized by the convolution of seismic wavelet and stratigraphic reflection coefficient.

[0053] S3: Neural network construction and parameter training. Construct a neural network model that includes a temporal feature extraction module and a spatiotemporal feature fusion module, and construct a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model.

[0054] S4: Neural network test. Calculate the envelope of the seismic data and use the seismic data, initial model, and seismic envelope as input to form a three-channel data set from the well-side i-channel. Similarly, standardize the data to be predicted and use the processed data to conduct neural network tests.

[0055] In Example 1, the elastic transformation process consists of three steps: first, a new spatial coordinate is obtained based on the original data coordinate space through affine transformation; then, the new coordinate space is smoothed using a Gaussian function to obtain a random displacement field; finally, the original data is mapped to the new coordinate space to obtain the perturbed data. The elastic transformation process is represented in the following form:

[0056]

[0057] Where D represents the generated random displacement field, T represents the time series, α controls the perturbation degree of the random displacement field, and δ is the Gaussian filter kernel size, controlling the smoothness of the random displacement field. The larger α and the smaller δ, the more severe the spatial perturbation; conversely, the greater the α and the smaller the δ, the smoother the spatial perturbation. The elastic transformation algorithm alters the distribution and thickness of reservoir parameters in the original sample, increasing the sample diversity. Figure 4 As shown, Figure 4 The results of different combinations of α and δ perturbations in the data augmentation method based on elastic transformation are shown below. Figure 4 In (a), α = 1300, δ = 200. Figure 4 In (b), α = 1300, δ = 130. Figure 4 In (c), α = 2000, δ = 130. Figure 4In (d), α = 130 and δ = 30. In this invention, α = 1300 and δ = 130 are selected, and the number of elastic transformations is set to 20. In addition to the original wave impedance model, 20 extended wave impedance models that can characterize the features of the original model and have more diversity are obtained through this process.

[0058] In Example 1, the seismic record is represented by the convolution of the seismic wavelet and the formation reflection coefficient. The relationship between wave impedance and reflection coefficient is as follows:

[0059]

[0060] Among them, t n This represents the nth time sampling point, Δt represents the time increment, R is the reflection coefficient, and Z is the wave impedance;

[0061] The convolution model process is represented as follows:

[0062] S=R*W+N (3)

[0063] Where S represents seismic data, R represents the reflection coefficient, W represents the Ricker wavelet, and N represents noise. The expression for the Ricker wavelet W is as follows:

[0064]

[0065] Where f m The dominant frequency of the seismic wavelet is denoted by t, which represents time. To further expand the diversity of the sample, the dominant frequency of the seismic data is determined by time-frequency analysis. Five dominant frequency information are obtained by adding or subtracting from the dominant frequency of the actual seismic data at 5 Hz intervals.

[0066] Preferably, five dominant frequency information points are obtained by adding or subtracting from the dominant frequency of the actual seismic data at 5Hz intervals, and the seismic record is then subjected to Hilbert transform using the following formula:

[0067]

[0068] Where y(t) represents the earthquake record, y H (t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

[0069] In Example 1, a high-pass filter is used to filter out the low-frequency components of the synthesized data according to the original data frequency band, simulating the lack of low-frequency components in actual seismic data. The envelope of the band-limited seismic data still carries long-wavelength information of the subsurface medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by calculating the amplitude of the analytic signal, and the process is written as follows:

[0070]

[0071] Where e represents the envelope of the seismic signal, the final sample library contains seismic data, seismic envelope, initial model and low-frequency model. Seismic data, seismic envelope and initial model are used as input, and low-frequency model is used as training label. The input data is taken from 5 well-side traces to form 101 pairs of two-dimensional data to enhance the spatial representation ability of the prediction results. The label data is one-dimensional data. The data is standardized to reduce the difficulty of neural network training. The low-frequency model is standardized using the mean and variance of the initial model.

[0072] In Example 1, as Figure 2 As shown in (a), a single gated recurrent unit (GRU) improves a simple recurrent neural network by coupling a reset gate and an update gate, and the process is represented as follows:

[0073]

[0074] Among them, u t r t X t , and h t Let represent the updated gate vector, reset gate vector, input sample, candidate hidden state, and hidden state at time t, respectively. w and b represent the weight parameters and bias terms, respectively. Sigmoid and tanh represent different activation functions. * represents the dot product operator.

[0075] In Example 1, as Figure 2 As shown in (b), the Bi-Gated Recurrent Network (Bi-GRU) unit builds a backpropagation layer on top of the GRU, taking into account both forward and backward time dependencies. The temporal feature extraction module consists of multiple parallel 3-in-1 Bi-GRUs. The number of parallel blocks depends on the value of i in step S4. If it is set to 2, there are 5 parallel blocks. Figure 2 (c) The spatiotemporal feature fusion module consists of a Bi-GRU and a fully connected layer.

[0076] In Example 1, the parameter training steps of the neural network include: initializing the network weight parameters w and bias term b; inputting the two-dimensional seismic data, initial model, and seismic envelope from the sample library into the neural network; outputting a one-dimensional predicted low-frequency model; and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows:

[0077]

[0078] Where N is the number of sampling points per channel, f is the network model, and Z is the number of sampling points per channel. L This is a practical low-frequency model, where S represents the seismic record and Z represents the actual low-frequency model. initFor the initial model, e is the earthquake envelope, W and b represent the network weights and biases, respectively. The first term of the loss function is the low-frequency wave impedance model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

[0079] In Example 1, the derivatives of the input data at each step of the neural network are recorded. The partial derivatives of the loss function with respect to the weight parameters and bias terms are obtained using the chain rule. Then, the Adam optimizer based on variable step size gradient descent is used to calculate the update amount of the weight parameters and bias terms. The learning rate is set to 0.05. The weight parameters and bias terms are updated. This process is repeated until the error of the loss function is minimized. Finally, the network weight parameters and bias terms are saved.

[0080] The low-frequency model construction method based on spatiotemporally gated recurrent unit fusion network disclosed in this invention also includes a neural network testing method, comprising the following steps:

[0081] (1) When constructing a low-frequency model for test data, first calculate the envelope of the seismic data and take the seismic data, the initial model and the seismic envelope as well-side i-channel data to form three channels as input. Similarly, the data to be predicted is standardized.

[0082] (2) Load the trained network weights and biases to construct a low-frequency model on the test data. Extract the variance σ and mean μ of the initial model to perform inverse standardization on the predicted low-frequency model. The inverse standardization is represented as follows:

[0083] x=σX+μ (9)

[0084] Where X represents standardized data and x represents destandardized data.

[0085] In Example 2, as Figure 3 The second embodiment shown utilizes the raw data from Marmousi2, where, Figure 3 (a) is the wave impedance. Figure 3 (b) is a synthetic seismic record. Figure 3 (c) is a low-frequency model. Figure 4 (d) is the initial model, which has a total of 2266 channels, each with 406 sampling points, and a time sampling rate of 6ms. Figure 3 The dashed lines in the text represent the selected pseudo-well data.

[0086] In Example 2, the low-frequency model construction method based on spatiotemporally gated recurrent unit fusion network (ST-GRU) of the present invention is compared with three neural network models with distinct characteristics. The comparison models include CNN-GRU, TCN, and UNet networks. CNN-GRU uses GRU as the basic unit while adding dilated convolutional layers to extract features at different scales. TCN uses dilated convolution as the basic unit, and different dilation rates obtain different receptive fields. UNet, as one of the most classic network frameworks, gradually increases the receptive field through a downsampling and upsampling encoder-decoder structure. Figure 5 The performance of different network models on the marmousi2 dataset is shown. Figure 5 (a) Figure 5 (c) Figure 5 (e) and Figure 5 (g) represents the prediction results of CNN-GRU, TCN, UNet, and ST-GRU. Figure 5 (b) Figure 5 (d) Figure 5 (f) and Figure 5 (h) represents the absolute error of predictions made by CNN-GRU, TCN, UNet, and ST-GRU, respectively. Figure 5 The results show that the low-frequency model construction method based on spatiotemporally gated cyclic unit fusion network (ST-GRU) has good spatial continuity, complete structural representation, inconspicuous background noise, and relatively clear characterization of the complex structure in the middle.

[0087] The above description only describes the present invention and its embodiments. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual structure is not limited to this. Therefore, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network, characterized in that, Includes the following steps: S1: Conduct wave impedance data augmentation. Based on the original dataset, introduce an elastic transformation algorithm for data augmentation. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through well logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the spatial disturbance in the longitudinal direction. S2: Establish a low-frequency model sample library. Seismic records are characterized by the convolution of seismic wavelet and stratigraphic reflection coefficient. S3: Neural network construction and parameter training. Construct a neural network model that includes a temporal feature extraction module and a spatiotemporal feature fusion module. The temporal feature extraction module consists of multiple parallel three-in-a-row bidirectional gated recurrent networks. The spatiotemporal feature fusion module consists of a bidirectional gated recurrent network and a fully connected layer. Construct a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. S4: Neural network test. Calculate the envelope of the seismic data and use the seismic data, initial model, and seismic envelope as input to form a three-channel data set from the well-side i-channel. Similarly, standardize the data to be predicted and use the processed data to conduct neural network tests.

2. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 1, characterized in that: The elastic transformation process consists of three steps. First, a new spatial coordinate system is obtained based on the original data coordinate space through an affine transformation. Then, a random displacement field is obtained by smoothing the new coordinate space using a Gaussian function. Finally, the original data is mapped to the new coordinate space to obtain the perturbed data. The elastic transformation process is represented in the following form: Where D represents the generated random displacement field, T represents the time series, α controls the perturbation degree of the random displacement field, and δ is the size of the Gaussian filter kernel, which controls the smoothness of the random displacement field.

3. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 1, characterized in that: Seismic records are represented by the convolution of the seismic wavelet and the stratigraphic reflection coefficient. The relationship between wave impedance and reflection coefficient is as follows: Among them, t n This represents the nth time sampling point, Δt represents the time increment, R is the reflection coefficient, and Z is the wave impedance; The convolution model process is represented as follows: S=R*W+N (3) Where S represents seismic data, R represents the reflection coefficient, W represents the Ricker wavelet, and N represents noise. The expression for the Ricker wavelet W is as follows: Where f m t represents the dominant frequency of the seismic wavelet.

4. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 3, characterized in that: Five dominant frequency information points were obtained by adding or subtracting from the dominant frequency of actual seismic data at 5Hz intervals. The Hilbert transform of the seismic record was then performed using the following formula: Where y(t) represents the earthquake record, y H (t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

5. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 4, characterized in that: Based on the original data frequency band, a high-pass filter is used to filter out the low-frequency components of the synthesized data to simulate the absence of low-frequency components in actual seismic data. The envelope of band-limited seismic data still carries long-wavelength information of the subsurface medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by calculating the amplitude of the analytic signal. The process is written as follows: Where e represents the envelope of the seismic signal.

6. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 1, characterized in that: A single gated recurrent unit improves a simple recurrent neural network by coupling a reset gate and an update gate, as shown in the following process: Among them, u t r t X t , and h t Let represent the updated gate vector, reset gate vector, input sample, candidate hidden state, and hidden state at time t, respectively. w and b represent the weight parameters and bias terms, respectively. Sigmoid and tanh represent different activation functions, and * represents the dot product operator.

7. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 6, characterized in that: The bidirectional gated recurrent network unit builds a backpropagation layer on the basis of a single gated recurrent unit, taking into account both forward and reverse time dependencies. The temporal feature extraction module consists of multiple parallel three-in-series bidirectional gated recurrent networks. The number of parallel blocks depends on the value of i in step S4. If it is set to 2, there are 5 parallel blocks. The spatiotemporal feature fusion module consists of a bidirectional gated recurrent network and a fully connected layer.

8. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 1, characterized in that: The parameter training steps of the neural network include: initializing the network weight parameters w and bias term b; inputting the two-dimensional seismic data, initial model, and seismic envelope from the sample library into the neural network; outputting a one-dimensional predicted low-frequency model; and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows: Where N is the number of sampling points per channel, f is the network model, and Z is the number of sampling points per channel. L This is a practical low-frequency model, where S represents the seismic record and Z represents the actual low-frequency model. init For the initial model, e is the earthquake envelope, W and b represent the network weights and biases, respectively. The first term of the loss function is the low-frequency wave impedance model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

9. The method for constructing a low-frequency model based on a spatiotemporally gated cyclic unit fusion network according to claim 8, characterized in that: The input data is recorded and the derivatives corresponding to each step of the neural network are recorded. The chain rule is used to obtain the partial derivatives of the loss function with respect to the weight parameters and bias terms. Then, the Adam optimizer based on variable step size gradient descent is used to calculate the update amount of the weight parameters and bias terms. The learning rate is set to 0.05, and the weight parameters and bias terms are updated. This process is repeated until the error of the loss function is minimized. Finally, the network weight parameters and bias terms are saved.

10. A method for constructing a low-frequency model based on a spatiotemporally gated recurrent unit fusion network according to any one of claims 1-9, further comprising a neural network testing method, characterized in that: Includes the following steps: (1) When constructing a low-frequency model for test data, first calculate the envelope of the seismic data and take the seismic data, the initial model and the seismic envelope from the well-side i-channel to form three channels as input. Similarly, the data to be predicted is standardized. (2) Load the trained network weights and biases to construct a low-frequency model on the test data. Extract the variance σ and mean μ of the initial model to perform inverse standardization on the predicted low-frequency model. The inverse standardization is represented as follows: x=σX+μ (9) Where X represents standardized data and x represents destandardized data.

Images