A three-dimensional pore pressure prediction method and device

Abstract

The present invention introduces a deep learning network model and seismic inversion results into three-dimensional pore pressure. Elastic parameter data such as density, longitudinal and transverse wave velocities of well logging are collected, as well as three-dimensional elastic attribute data obtained through seismic inversion as a data set. An optimization algorithm is used to optimize and train the network model. The trained model is applied to the test area to calculate and output the three-dimensional pore pressure prediction results. At the same time, multi-source data such as well logging and seismic data provide more comprehensive geological feature attributes, can jointly learn deep features from multi-source data, and better predict pore pressure.

Description

Technical Field

[0001] The present invention relates to a three-dimensional pore pressure prediction method and device, belonging to the technical field of oil and gas field development and coalfield exploration and mining. Background Art

[0002] Formation pore pressure is a key factor in evaluating reservoir stability, designing drilling programs, and predicting production dynamics. Pore pressure refers to the force exerted on the fluid within the pores of a rock formation, which is usually related to the formation depth and is expressed in static hydrostatic pressure, such as megapascals or pounds per square inch. Abnormal pore pressure can manifest as overpressure and underpressure. Overpressure occurs when the pore pressure exceeds the normal level due to additional pressure sources, which may trigger dangerous events such as well kicks. Underpressure means that the pore pressure is lower than the normal level, which may lead to problems such as well losses and circulation losses, increasing the drilling cost.

[0003] Current pore pressure prediction methods can be divided into direct methods and indirect methods. Direct methods include leak tests, rock mechanics experiments, and seismic data acquisition, aiming to directly measure the formation pressure. However, these methods usually can only provide measurements at specific depths, making it difficult to obtain continuous pore pressure profiles. Moreover, the process of collecting downhole data is costly and risky. Indirect methods usually rely on theoretical models to deduce pore pressure. Among these methods, the more common ones are the use of the undercompaction theory and the effective stress method.

[0004] In recent years, the rapid development of machine learning technology, especially the rise of deep learning, has provided new possibilities for the research of pore pressure prediction. Data-driven models offer a simpler establishment process and stronger adaptability compared to traditional mechanical models, and can handle various diverse, large, and complex data. These models are good at capturing the internal patterns of data and have strong generalization capabilities. Many studies have explored the application of machine learning in pore pressure prediction. Although machine learning models have achieved success in some applications, many existing models still face challenges in capturing the internal time dynamics of data. Existing models perform poorly in processing time series data, resulting in reduced prediction accuracy.

[0005] To address the deficiencies of traditional methods, the present invention proposes a new deep learning network model framework for predicting pore pressure. This model combines the efficient parallel computing ability of convolutional neural networks and the sequential processing advantages of recurrent neural networks, effectively identifying abnormal pressure events across multiple layers. Moreover, through a convolutional acceleration mechanism, the computing efficiency is improved, making it suitable for real-time prediction of large-scale three-dimensional data. At the same time, by combining deep learning technology with high-precision seismic inversion, not only can high-precision three-dimensional pore pressure prediction be achieved, but also the lateral variations of pore pressure that are difficult to detect by traditional methods can be revealed, providing strong technical support for the precise management and efficient development of oil and gas fields. Summary of the Invention

[0006] In view of the deficiencies in the prior art, the present invention provides a method and device for predicting pore pressure based on a deep learning network model to solve the deficiencies of the prior art.

[0007] In a first aspect, the present invention provides a three-dimensional pore pressure prediction method based on a deep learning network model, comprising the following steps:

[0008] S1: Collect well logging data and seismic data for preprocessing to construct a data set. The data set includes a training set and a validation set. The training set includes labeled training data and unlabeled training data. The labeled training data includes relevant curve data in well logging data, such as density, longitudinal wave velocity P, and shear wave velocity S. The unlabeled training data is three-dimensional elastic parameter information obtained by prestack inversion of seismic data;

[0009] Further, relevant curve data in well logging data, such as density, longitudinal wave velocity, and shear wave velocity, are extracted as labeled training data, and three-dimensional elastic parameter information obtained by prestack inversion of seismic data is used as unlabeled training data;

[0010] S2: Define the initial parameters and initial model of the model, establish the deep learning network model structure according to the discretized state space model framework, and use the collected well logging and seismic data for iterative training to ensure that the model learns the non-linear relationship in the input data;

[0011] S3: Randomly select training set data and validation set data in a certain proportion to train the model, and continuously iterate the model by adjusting the hyperparameters of the state size and sequence length until the value of the loss function falls within the preset threshold range or the number of iterations reaches the preset value;

[0012] S4: Input the three-dimensional density, longitudinal wave velocity P, and shear wave velocity S of the three-dimensional seismic data into the trained model, and output the prediction result of the model to obtain the prediction result of the three-dimensional pore pressure.

[0013] In a second aspect, the present invention also provides a three-dimensional pore pressure prediction device based on a deep learning network model, comprising:

[0014] A data acquisition module that collects well logging data and seismic data to construct a data set. The data set includes a training set, a validation set, and a test set. The training set includes labeled training data and unlabeled training data.

[0015] A model construction module that establishes the deep learning network model structure according to the discretized state space model framework, and uses the collected well logging and seismic data for iterative training to ensure that the model learns the non-linear relationship in the input data;

[0016] The model training module randomly selects training set data and validation set data to train the model until the value of the loss function falls within a preset threshold range or the number of iterations reaches a preset value;

[0017] Specifically, the training set data and validation set data can be randomly selected in a ratio of 8:2 for model training. The training process includes two parts: forward propagation and backpropagation parameter update. The model is continuously iterated by adjusting hyperparameters until the value of the loss function falls within a preset threshold range or the number of iterations reaches a preset value;

[0018] The prediction module inputs the three-dimensional density, longitudinal wave velocity P, and transverse wave velocity S of the three-dimensional seismic data into the trained model, outputs the prediction result of the model, and obtains the prediction result of the three-dimensional pore pressure.

[0019] The beneficial effects of the present invention are as follows: Existing machine learning methods often have difficulty capturing the internal temporal dynamics of geological data, such as well logging and seismic data, resulting in less than ideal prediction accuracy; Most existing pore pressure prediction techniques mainly focus on one-dimensional prediction and are not adapted to three-dimensional applications. The method proposed in this paper successfully realizes three-dimensional pore pressure prediction, highlighting the deficiencies of other methods in this regard. The present invention combines a deep learning network model with seismic inversion technology to jointly extract pore pressure features from multi-source data such as well logging and seismic data. The deep learning network model uses convolutional operations instead of recursive inference, reducing the computational complexity of processing long sequence data. Compared with traditional recursive operations, convolutional operations can be processed in parallel, thus significantly improving the computational efficiency. It has obvious advantages in parallel processing, especially in large-scale data sets, which can significantly reduce the training time. At the same time, the long-term dependencies and complex non-linear trends in different data can be effectively captured through the deep learning network model without scale matching, improving the prediction accuracy and reliability, as well as the model generalization and pore pressure prediction accuracy, and can accurately predict the lithofacies distribution in unknown wells or well-free areas. Description of the Drawings

[0020] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0021] Figure 1 It is a flowchart of a method for predicting three-dimensional pore pressure based on a deep learning network model according to an embodiment of the present invention.

[0022] Figure 2 It is the input attributes and label data of the labeled training data input according to an embodiment of the present invention.

[0023] Figure 3 It is the input attribute of the tagless seismic inversion elastic parameter training data input in an embodiment of the present invention.

[0024] Figure 4 They are the input data and prediction results in an embodiment of the present invention.

[0025] Figure 5 It is the three-dimensional pore pressure prediction result in an embodiment of the present invention. Detailed implementation manners

[0026] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0027] As Figure 1 shown, in the first aspect, the present invention provides a three-dimensional pore pressure prediction method, including the following steps:

[0028] S1: Collect logging data and seismic data for preprocessing to construct a data set, which includes a training set, a validation set, and the training set includes labeled training data and unlabeled training data;

[0029] Further, as Figures 2 - 3 shown, the labeled training data includes relevant curve data in logging data, such as density, longitudinal wave velocity P, and shear wave velocity S, and the unlabeled training data is three-dimensional elastic parameter information obtained through prestack inversion of seismic data;

[0030] Further, the preprocessing includes feature extraction and then normalization of the elastic parameters obtained by seismic inversion, and mapping them to a specific interval to accelerate the training and convergence speed of the model. The normalization formula is as follows:

[0031]

[0032] where d is the eigenvalue of the elastic parameter data, dmin and dmax respectively represent the minimum and maximum values in the data set, and d` is the value after normalization, with a range of [0,1].

[0033] S2: Define the initial parameters and initial model of the model, establish the deep learning network model structure according to the discretized state space model framework, and use the collected logging and seismic data for iterative training to ensure that the model learns the nonlinear relationship in the input data; The initial architecture of the deep learning network model is as follows:

[0034] Among them, h(t) represents the hidden state, describing the current internal state of the system, and h'(t) represents the next spatial state; x(t) represents the external input, affecting the behavior of the system; y(t) is the output of the system; where A(t) ∈ R N×N , B(t) ∈ R N×M , C(t) ∈ R M×N , and they are usually also called the state matrix, input matrix, and output matrix.

[0035] Define the initial parameters of the model, including the state matrix A, input matrix, output matrix, time step, and convolution kernel size;

[0036] Specifically, determine the initial state matrix A, as shown in the following formula,

[0037]

[0038] where, A nk represents the element in the nth row and kth column of the matrix; B n represents the element of the matrix, and they jointly describe the evolution of the state space and the approximation process of the function.

[0039] It is also possible to choose a diagonal matrix or a more complex structured matrix, such as a DPLR matrix. Set the initial values of the input matrix B and output matrix C, and select an appropriate time step for discretizing the model.

[0040] Construct a discrete-time model: Convert the continuous-time model into a discrete-time model to obtain the discrete-time state equation, as shown in the following formula;

[0041]

[0042] where, x k is the state vector at the discrete time point k, k represents the discrete time step, and as k increases, the system state x k will be updated to the state x k+1 at the next time step; f k is the input vector.

[0043] In order to make it applicable to pore pressure prediction, the initial architecture of deep learning needs to be discretized into u = (u0, u1, u2,...) so that deep learning can effectively process well logging or seismic input data. The discretized state transition equation is:

[0044]

[0045] y k = Ch k y krefers to the discretized output; h k represents the hidden state at time k-1, h k-1 represents the hidden state at the previous time;

[0046] Among them, represents the matrix after discretization, which can be expressed as:

[0047]

[0048] Among them, exp refers to the natural exponential function; (Δ, A, B) are continuous parameters, is a discrete parameter, and this transformation is defined by the formulas and where (f A , f B ) is the discretization rule to ensure that the model can maintain stable performance and accuracy when processing logging and seismic time series.

[0049] According to the discretized state equation, the convolution kernel is defined. The size of the convolution kernel is determined according to the resolution of the input data to ensure that the convolution kernel can capture long-range dependencies, so that the output y can be expressed as the convolution of the input x and the convolution kernel , that is, it can be expressed as:

[0050]

[0051] Among them the derivation of is based on the assumption that the initial state is x -1 = 0. Since the discrete recursion is linear, it is expanded analytically to obtain the equivalent convolution form of the model

[0052]

[0053] In other words, for the input x, the output y has a simple closed form:

[0054]

[0055] Then it can be vectorized into a single convolution through the explicit formula of the kernel:

[0056]

[0057] This equation is a single convolution, simply referred to as the state space kernel.

[0058] The convolution operation is calculated through the following state space kernel, and the input sequence is processed in parallel to improve the calculation speed:

[0059]

[0060] Among them, h i The weight of the state space kernel, u k-i is the input sequence. Among them, the deep learning network model enhances the flexibility and adaptability of the model by associating the weight parameters with the input data. Specifically, the parameters of the model can be dynamically adjusted according to the characteristics of the input, so as to achieve optimized prediction for different data.

[0061] Adopting a selective state space model enables the model parameters to be dynamically adjusted according to the characteristics of the input data

[0062] B = S B (x)

[0063] C = S C (x)

[0064] Δ = τA(Parameter + S A (x))(8)

[0065] In the above formula: S B (x), S C (x) represents a matrix dynamically generated according to the input x, allowing the model to adjust its behavior according to different characteristics of the input; Δ represents a learnable parameter used to estimate the discrete interval of the input samples, enhancing the processing of time dependence.

[0066] S3: Randomly select the training set data and validation set data according to a certain proportion to train the network model, and continuously iterate the model by adjusting the hyperparameters until the value of the loss function falls within the preset threshold range, or the number of iterations reaches the preset value;

[0067] Specifically, the model training method in step S3 specifically includes the following content: considering the influence of different input attributes on the prediction results, inputting seismic attributes and logging parameters into the network model to obtain the prediction results, calculating the error between the output value of the output layer and the actual value, and adjusting the model hyperparameters according to the loss error, such as the learning rate of the model, the number of network layers, the number of neurons, the state size, and the sequence length. Repeat the above steps to optimize the deep learning network model multiple times until the value of the loss function falls within the preset threshold range, or the number of iterations reaches the preset value;

[0068] Specifically, randomly select the training set data and validation set data according to the ratio of 8:2 for model training. The training process of the neural network includes two parts: forward propagation and backpropagation parameter update. In the forward propagation process, the output signal of each layer will be transmitted to the next layer as the input signal. These input signals pass through the network model and the additionally added hidden layers in turn, and finally reach the output layer, complete the non-linear transformation through the activation function, and output the prediction results;

[0069] Loss function calculation:

[0070] Among them, L sz is the cross-entropy loss function, y i is the soft label, P i is the predicted probability distribution of the pore pressure by the trained model, n is the number of parameter categories, T is the temperature factor, the adjustment factor of the cross-entropy loss function, is the sum of absolute errors, representing the average deviation between the predicted probability distribution of the model and the soft label produced by the adjusted network model. max(y) is the maximum value of the soft label, representing the maximum probability value output by the adjusted network model among all categories. min(y) is the minimum value of the soft label, representing the minimum probability value output by the adjusted processing network model among all categories.

[0071] Specifically, the network model architecture in an embodiment of the present invention is multiple convolutional layers, a linear projection layer, and a dynamic state space module. In this model, the convolutional unit consists of two convolutional layers. The first convolutional layer contains 32 3x3 convolutional kernels, with a stride of 1, and uses the SiLU activation function; the second convolutional layer contains 64 3x3 convolutional kernels, also with a stride of 1, and also uses the SiLU activation function; in the dynamic state space module, the number of states is set to 64, and the sequence length is set to 100;

[0072] S4: Input the three-dimensional density, P-wave velocity, and S-wave velocity of the three-dimensional seismic data into the trained deep learning network model, output the prediction result of the model, and obtain the prediction result of the three-dimensional pore pressure.

[0073] As Figure 4 shown, it is a schematic diagram of the input data and prediction result in an embodiment of the present invention. Figure 4 a to Figure 4 c are the P-wave velocity, S-wave velocity, density, and prediction result. Figure 4 d is the true logging pore pressure (solid line) and the prediction result (dashed line).

[0074] As Figure 5 shown, it is the prediction result of applying the model to the three-dimensional pore pressure in the area of the three-dimensional seismic data to be predicted in an embodiment of the present invention. All wells are included in the training process, and data such as density, P-wave, and S-wave velocity are input into the optimized deep learning network model to obtain the prediction result of the three-dimensional pore pressure.

[0075] The present invention introduces the deep learning network model and seismic inversion results into the three-dimensional pore pressure. At the same time, using multi-source data such as logging and seismic provides more comprehensive geological feature attributes, can jointly learn deep features from multi-source data, and better predict the pore pressure. Due to the above technical solutions adopted by the present invention, it has the following advantages:

[0076] 1. Extract features such as density and wave velocity from well logging data, and extract elastic property features from seismic data, providing rich input features for pore pressure prediction and helping to improve the prediction ability of the model;

[0077] 2. Through the deep learning network model of the present invention, comprehensively utilize labeled well logging data and unlabeled seismic data for training, providing rich input features for pore pressure prediction and helping to improve the prediction ability of the model;

[0078] 3. Break through the one-dimensional limitation of traditional methods, achieve high-precision prediction of three-dimensional pore pressure, and the prediction results have a high degree of matching with the actual measured values and smaller errors;

[0079] 4. The deep learning network model combines the efficient parallel computing ability of the convolutional neural network and the sequential processing advantage of the recurrent neural network, improves the computing efficiency, and can process large-scale three-dimensional data in real time.

[0080] In a second aspect, the present invention also provides a three-dimensional pore pressure prediction device based on a deep learning network model, including:

[0081] A data acquisition module that acquires well logging data and seismic data for preprocessing, constructs a data set, the data set includes a training set and a validation set, the training set includes labeled training data and unlabeled training data, the labeled training data includes relevant curve data in well logging data, such as density, longitudinal wave velocity P, and transverse wave velocity S, and the unlabeled training data is three-dimensional elastic parameter information obtained through prestack inversion of seismic data;

[0082] A model construction module that defines the initial parameters and the initial model of the model, establishes the deep learning network model structure according to the discretized state space model framework, and uses the acquired well logging and seismic data for iterative training to ensure that the model learns the non-linear relationship in the input data;

[0083] A model training module that randomly selects training set data and validation set data in a certain proportion to train and validate the model, and continuously iterates the model by adjusting the state size and sequence length hyperparameters until the value of the loss function falls within the preset threshold range or the number of iterations reaches the preset value;

[0084] A prediction module that inputs the three-dimensional density, longitudinal wave velocity P, and transverse wave velocity S of the three-dimensional seismic data into the trained deep learning network model, outputs the prediction result of the model, and obtains the prediction result of the three-dimensional pore pressure.

[0085] The present invention provides a computer storage medium, which stores at least one instruction adapted to be loaded and executed by a processor to perform the method steps of one or more embodiments of this specification.

[0086] The present invention provides an electronic device, which may include: a processor and a memory; wherein, the memory stores a computer program adapted to be loaded and executed by the processor to perform the method steps of one or more embodiments of this specification.

[0087] The above embodiments only represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A three-dimensional pore pressure prediction method based on a deep learning network model, characterized in that, It includes the following steps: S1: Collect well logging data and seismic data for preprocessing, and construct a data set. The data set includes a training set and a validation set. The training set includes labeled training data and unlabeled training data. The labeled training data includes relevant curve data in well logging data, such as density, longitudinal wave velocity P, and shear wave velocity S. The unlabeled training data is three-dimensional elastic parameter information obtained by prestack inversion of seismic data; S2: Define the initial parameters and initial model of the model. Establish the deep learning network model structure according to the discretized state space model framework, and use the collected well logging and seismic data for iterative training to ensure that the model learns the non-linear relationship in the input data; Defining the initial parameters of the model includes the state matrix A, input matrix B, output matrix C, time step, and convolution kernel size. The initial architecture of the deep learning network model is as follows h'(t) = A(t)x(t) + B(t)u(t) y(t) = C(t)x(t) + D(t)u(t) Among them, h(t) represents the hidden state, describing the current internal state of the system, h'(t) represents the next spatial state; x(t) represents the external input, which affects the behavior of the system; y(t) is the output of the system; where A(t) ∈ R N×N , B(t) ∈ R N×M , C(t) ∈ R M×N ; Specifically, determine the initial state matrix A, as shown in the following formula, Among them, A nk represents the element in the n-th row and k-th column of the matrix; B n represents the -th element of the matrix; Construct a discrete-time model: Convert the continuous-time model into a discrete-time model to obtain the discrete-time state equation, as shown in the following formula; where x k is the state vector at discrete time point k, where k represents the discrete time step, and as k increases, the system state x k is updated to the state x k+1 at the next time step; f k is the input vector; S3: Randomly select training set data and validation set data in a certain proportion to train and validate the model. Continuously iterate the model by adjusting the state size and sequence length hyperparameters until the value of the loss function falls within the preset threshold range or the number of iterations reaches the preset value; S4: Input the three-dimensional density, longitudinal wave velocity P, and shear wave velocity S of the three-dimensional seismic data into the trained deep learning network model, output the prediction result of the model, and obtain the prediction result of the three-dimensional pore pressure.

2. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 1, wherein: The preprocessing includes feature extraction and then normalization of the elastic parameters obtained by seismic inversion, and mapping them to a specific interval to accelerate the training and convergence speed of the model. The normalization formula is as follows: Where d is the eigenvalue of the elastic parameter data, dmin and dmax respectively represent the minimum and maximum values in the data set, and d` is the value after normalization, with a range of [0,1].

3. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 1, characterized in that: Step S1 is specifically to randomly divide the data into a training set and a validation set in a ratio of 8:

2. The training process includes two parts: forward propagation and backpropagation parameter update.

4. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 3, wherein, Backpropagation and parameter update include applying the trained model to the test well, inputting the corresponding attributes, calculating the error between the output value of the output layer and the actual value, and adjusting the model hyperparameters according to the loss error, such as the learning rate of the model, the number of network layers, and the number of neurons; Loss function calculation: Among them, L sz is the cross-entropy loss function, y i is the soft label, P i is the predicted probability distribution of the trained model for pore pressure, n is the number of parameter categories, T is the temperature factor, is the adjustment factor of the cross-entropy loss function, is the sum of absolute errors, representing the average deviation between the predicted probability distribution of the model and the soft label produced by the adjusted network model. max(y) is the maximum value of the soft label, representing the maximum probability value output by the adjusted network model among all categories. min(y) is the minimum value of the soft label, representing the minimum probability value output by the adjusted processing network model among all categories.

5. The three-dimensional pore pressure prediction method based on the deep learning network model according to claim 1, characterized in that: Discretize the initial architecture of deep learning into u = (u0, u1, u2,...) so that deep learning can effectively process input data from well logging or seismic. The discretized state transition equation is: y t = Ch(t) y t refers to the discretized output, where t specifically refers to the time point after discretization; h(t) represents the hidden state at time t, and h(t - 1) represents the hidden state at the previous time point; The following formula is used to calculate the discrete matrix during the discretization process: Among them, represents the discretized matrix, which converts the continuous parameters (Δ, A, B) into discrete parameters This transformation is defined by the formulas and where (f A , f B ) is the discretization rule; Define the convolution kernel according to the discretized state equation The size of the convolution kernel is determined according to the resolution of the input data to ensure that the convolution kernel can capture long-range dependencies, so that the output y can be expressed as the convolution of the input x and the convolution kernel That is, it can be expressed as: Assume that the initial state is x -1 = 0. Since the discrete recursion is linear, expand it analytically to obtain the equivalent convolutional form of the model In other words, for the input x, the output y has a simple closed form: Then it can be vectorized into a single convolution through the explicit formula of the kernel: This equation is a single convolution, simply referred to as the state space kernel.

6. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 5, wherein, Adopt a selective state space model so that the model parameters can be dynamically adjusted according to the characteristics of the input data; B = S B (x) C=S C (x) Δ = τA(Parameter + S A (x)) In the above formula: S B (x), S C (x) represents a matrix dynamically generated according to the input x, allowing the model to adjust its behavior according to different features of the input; Δ represents a learnable parameter used to estimate the discrete interval of input samples, enhancing the processing of time dependence.

7. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 6, characterized in that: The convolution operation is calculated through the following state-space kernel, processing the input sequence in parallel to improve the calculation speed: where h i is the weight of the state space kernel, and u k-i is the input sequence.

8. The three-dimensional pore pressure prediction method based on a deep learning network model according to claim 7, wherein: The network model architecture consists of multiple convolutional layers, linear projection layers, and dynamic state-space modules. In this model, the convolutional unit is composed of two convolutional layers. The first convolutional layer contains 32 3x3 convolutional kernels with a stride of 1 and uses the SiLU activation function; the second convolutional layer contains 64 3x3 convolutional kernels with a stride of 1 and also uses the SiLU activation function; in the dynamic state-space module, the number of states is set to 64 and the sequence length is set to 100.

9. A three-dimensional pore pressure prediction device based on a deep learning network model, comprising: A data acquisition module that acquires well logging data and seismic data for preprocessing, constructs a data set. The data set includes a training set and a validation set. The training set includes labeled training data and unlabeled training data. The labeled training data includes relevant curve data in well logging data, such as density, longitudinal wave velocity P, and transverse wave velocity S. The unlabeled training data is three-dimensional elastic parameter information obtained through prestack inversion of seismic data; A model construction module that defines the initial parameters and initial model of the model, establishes the deep learning network model structure according to the discretized state-space model framework, and uses the acquired well logging and seismic data for iterative training to ensure that the model learns the non-linear relationship in the input data; Defining the initial parameters of the model includes the state matrix A, input matrix B, output matrix C, time step, and convolutional kernel size. The initial architecture of the deep learning network model is as follows: h'(t) = A(t)x(t) + B(t)u(t) y(t) = C(t)x(t) + D(t)u(t) Among them, h(t) represents the hidden state, describing the current internal state of the system, and h'(t) represents the next spatial state; x(t) represents the external input, affecting the behavior of the system; y(t) is the output of the system; where A(t) ∈ R N×N , B(t) ∈ R N×M , C(t) ∈ R M×N ; Specifically, determine the initial state matrix A as shown in the following formula: Among them, A nk represents the element in the n-th row and k-th column of the matrix; B n represents the th element of the matrix; Construct a discrete-time model: Convert the continuous-time model into a discrete-time model to obtain the discrete-time state equation as shown in the following formula; where, x k is the state vector at discrete time point k, where k represents the discrete time step number. As k increases, the system state x k is updated to the state x k+1 at the next time step; f k is the input vector; A model training module that randomly selects training set data and validation set data in a certain proportion to train and validate the model, and continuously iterates the model by adjusting the hyperparameters of the state size and sequence length until the value of the loss function falls within the preset threshold range or the number of iterations reaches the preset value; A prediction module that inputs the three-dimensional density, longitudinal wave velocity P, and transverse wave velocity S of the three-dimensional seismic data into the trained deep learning network model, outputs the prediction result of the model, and obtains the prediction result of the three-dimensional pore pressure.

Images