Method for predicting under-compaction low-speed mudstone distribution based on geometric factor parameters

Abstract

The invention discloses a method for predicting under-compaction low-speed mudstone distribution based on geometric factor parameters, and belongs to the field of geological engineering and exploration. An LSTM multilayer long short-term memory network structure is set, and training is performed through a preset training set and a preset test set; iteratively calculating the incidence relation of the target function and the geometric factor data set in the time dimension by adopting a gradient updating algorithm of time truncation back propagation, completing the training of the LSTM network, and obtaining a geometric factor prediction model; and evaluating the performance of the model by using a mean absolute error MAE and a root-mean-square error RMS, and finally predicting a geometric factor according to the LSTM network obtained through optimization and determining the distribution range of the under-compacted mudstone and the compaction degree of the under-compacted mudstone. Therefore, the possible development range of the low-speed mudstone is identified, so that the probability of falling into a'bright spot 'trap is reduced before exploration and development.

Description

Technical Field

[0001] The invention belongs to the field of geological engineering and exploration technology, and in particular relates to a method for predicting the distribution of under-compacted low-speed mudstone based on geometric factor parameters. Background Art

[0002] The Yinggehai Basin is located between the Indochina Block and the South China Block. It is a large NW-trending Cenozoic strike-slip extensional basin formed by multiple factors such as the Indo-Tibetan continental collision and the seafloor expansion of the South China Sea. It is rich in natural gas resources and is one of the main battlefields for natural gas exploration in the western South China Sea. In recent years, the Yinggehai Basin has often encountered low-velocity mudstone layers with "bright spot" seismic response characteristics during the development process. This is consistent with the traditional understanding of the characteristics of gas-bearing reservoirs, and therefore often brings traps to the identification of the target layer in exploration and development.

[0003] A large number of scholars have studied the geology, geophysics, logging, well logging and core data of low-velocity mudstone wells and divided the causes of low-velocity mudstone into three types: undercompacted, gas-bearing and fractured. The Yinggehai Basin is a basin of rapid subsidence and rapid sedimentation. Undercompacted strata are common in the strata below the Huangliu Formation under the protection of overpressure, and undercompacted mudstone is widely developed. Summary of the invention

[0004] In order to solve the technical problems existing in the background technology, the present invention aims to provide a method for predicting the distribution of under-compacted low-speed mudstone based on geometric factor parameters. The distribution of geometric factors in the target area is calculated through deep learning of a multi-layer long short-term memory network. The different range values ​​of the mudstone geometric factor represent the degree of compaction of the mudstone. Combined with the cause of under-compacted low-speed mudstone and analyzed in combination with geological data, the possibility of drilling low-speed mudstone is reduced in advance before exploration and development.

[0005] In order to solve the technical problem, the technical solution of the present invention is:

[0006] A method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters, the method comprising:

[0007] Set up the LSTM multi-layer long short-term memory network structure, train it with the preset training set and test set, use the time truncated back propagation gradient update algorithm to iteratively calculate the correlation between the objective function and the geometric factor data set in the time dimension, complete the training of the LSTM network, and obtain the geometric factor prediction model;

[0008] The mean absolute error (MAE) and root mean square error (RMS) were used to evaluate the model performance. Finally, the optimized LSTM network was used to predict the geometric factors and determine the distribution range of undercompacted mudstone and its compaction degree.

[0009] Furthermore, the LSTM multi-layer long short-term memory network structure is set, specifically including:

[0010] The first layer is the forget gate, which is used to select and retain the previous step time C t-1 and the current step time C t The memory information of the forget gate is expressed as:

[0011] f t =σ(w f [s t-1 ,x t ]+b f )

[0012] In the formula, w f is the weight coefficient matrix of the forget gate layer, b f is the offset vector of the forget gate layer, w fs and w fx For input information t-1 and x t The corresponding weight coefficient matrix, f t is the forget gate at the current time t;

[0013] The second layer is the input gate structure; the input gate controls the transmission of information to the cell state C and processes the input of the current sequence data; this layer consists of two parts, the first part uses the sigmoid function to determine the updated information state, and the second part uses the tanh unit to calculate a new vector:

[0014] i t =σ(w i [s t-1 ,x t ]+b i )

[0015]

[0016] where i t is the input gate, σ(·) is the sigmoid function, w i and b i are the weight coefficient matrix and offset vector of the input gate, respectively, t is the current cell state of the input layer;

[0017] The third layer is the cell state of unit C; the addition and deletion of information are determined by the forget gate and input gate; the cell state of the previous step is multiplied by the forget gate f t , determine C t-1 The information that needs to be discarded is added to the information of the input gate to perform the final cell update:

[0018]

[0019] The last layer is the output gate structure. The sigmoid function is used to determine the cell unit output, the tanh activation function is used to process the cell unit, and the sigmoid gate is used to calculate the output:

[0020] O t =σ(w o [s t-1 ,x t ]+b o )

[0021] S t =O t ⊙tanh(C t )

[0022] Where S t is the hidden state output, σ(·) is the sigmoid function, w o and b o is the weight matrix and the offset vector of the output gate.

[0023] Furthermore, the elastic parameters and logging data are organized into training sets and test sets, where the data used as input features include porosity, permeability, water saturation, P-wave velocity and S-wave velocity; the data used as target labels are geometric factors obtained through modeling; the data are divided into training data sets and test data sets to evaluate the model performance.

[0024] Furthermore, the gradient update algorithm of time-truncated back-propagation is used to iteratively calculate the relationship between the objective function and the geometric factor data set in the time dimension, including:

[0025] The output value of each neuron in the LSTM forward propagation is f t 、i t , C t , O t , S t ; In LSTM, the propagation of the error term δ through reverse calculation includes the following directions: one is the back propagation over time; the other is the error term propagated to the previous layer; the gradient of each weight is calculated using the corresponding error term.

[0026] Furthermore, the mean absolute error (MAE) and root mean square error (RMS) are used to evaluate the model performance, including:

[0027] Define the original input parameters as X0 = {X1, X2, ..., X n}, training set X train ={X1,X2,…,X r} and test set X test ={X1,X2,…,X e}Divided into 7:3;

[0028] The mean absolute error (MAE) and root mean square error (RMS error) are used to estimate the performance of the model:

[0029]

[0030] Among them, y i and p i are the actual geometric factor and the predicted geometric factor respectively, and N is the number of samples.

[0031] Furthermore, after training, the LSTM network can predict the target area through the learned model relationship and predict the corresponding geometric factors; combined with geological characteristics and sedimentary environment, it can identify the distribution and maturity of undercompacted low-speed mudstone.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] Enhanced prediction capabilities: Deep learning is used to effectively simulate the causes and distribution of undercompacted mudstone in geological processes, thereby improving the ability to identify the degree of compaction of mudstone.

[0034] Reduce exploration risks: It can identify possible development areas of low-velocity mudstone in advance, play an important supporting role in subsequent exploration and development, and reduce drilling risks.

[0035] Practicality and innovation: Combining traditional geological models with modern deep learning methods, the method highlights its innovation and has practical application value in the geological exploration industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 , main flow chart of the method for predicting the distribution of under-compacted low-speed mudstone based on geometric factor parameters of the present invention. DETAILED DESCRIPTION

[0037] The specific implementation mode of the present invention is described below in conjunction with embodiments:

[0038] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0039] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0040] Embodiment 1:

[0041] The mudstone with a particularly good degree of compaction is called shale. An existing shale anisotropic high temperature and high pressure rock physics modeling method is used, which includes the following steps in sequence:

[0042] Step 1: Using the Voigt-Reuss-Hill average method (one of the synthesis methods), clay, quartz and calcite are synthesized into shale according to the elastic parameters and volume percentage of each mineral.

[0043] The Voigt average is the upper limit of the limit and can also be called the equal strain average. It assumes that each component has the same strain and further gives the ratio of the average stress to the average strain. The formula is:

[0044]

[0045] Where M V ——represents the equivalent elastic modulus of N components of rock;

[0046] M i ——represents the elastic modulus of the ith component (preferably shear modulus and bulk modulus);

[0047] f i ——represents the volume content of the ith component.

[0048] The Reuss average is the lower limit of the limit, and can also be called the equal stress average. It assumes that each component has the same strain and further gives the ratio of the average stress to the average strain. The formula is:

[0049]

[0050] The meanings of the parameters in the formula are the same as those in the Voigt formula. Hill uses the arithmetic average of the Voigt upper limit and the Reuss lower limit to estimate the equivalent elastic modulus of rock, and VRH is expressed as:

[0051]

[0052] The elastic parameters and volume percentage of each mineral can be obtained by X-ray Diffraction analysis (XRD) of shale samples.

[0053] Step 2: Use the porosity measured by well logging to obtain the geometric factors of the crystal orientation function.

[0054] The crystal orientation distribution function is represented by w(ξ,ψ,φ), which satisfies:

[0055]

[0056] Roe (1964) expanded the above equation into a series of spherical harmonic functions by processing a set of X-ray diffraction data.

[0057]

[0058] Inversely, the geometric factor of the crystal orientation function is obtained:

[0059]

[0060] Wherein, ξ, ψ, φ respectively represent the angles between the xoy plane and the XOY plane, the xoz plane and the XOZ plane, and the yoz plane and the YOZ plane formed by two coordinate axis systems when the mudstone matrix is ​​distributed in three-dimensional space, and the two coordinate axis systems include: a space coordinate axis system (x, o, y) with the positive up, down, front, back, left, and right of the space as the coordinate axes, and an object coordinate axis system (X, O, Y) with the positive up, down, front, back, left, and right of the matrix object in the space as the coordinate axes, ξ, ψ, φ represent the directional distribution function of the mudstone matrix, Wlmn represents the Legendre coefficient corresponding to the spherical harmonic function, Zlmn(ξ) represents the spherical harmonic function based on the ξ angle expansion, e represents the natural exponent, i represents the imaginary unit, and l, m, n are the Legendre coefficients corresponding to the expansion function.

[0061] The matrix direction distribution function under mudstone compaction is as follows:

[0062]

[0063] Wherein, W(ξ) represents the distribution function based on the ξ angle expansion, and A represents the geometric parameters of the mudstone compaction.

[0064] In addition, Ran Bachrach (2010) proposed that there is a certain relationship between the compaction orientation distribution function or other statistical orientation distribution functions and the pore shape and arrangement. lmn The corresponding model parameter m lmn . This simplifies the W based on the porosity φ lmn The process of seeking;

[0065]

[0066] where φ c is the critical porosity. The maximum value is obtained in the case of perfect alignment, n200 and m 400 is a normalized model parameter based on the pore aspect ratio and compaction direction distribution function. For the compaction model n 200 =1,m 400 =2.

[0067] Step 3: Substitute the obtained geometric factors of the crystal orientation function into the matrix model to obtain the shale anisotropic matrix model.

[0068] Sayers (1995) proposed to use the orthogonal relationship between shale distribution functions (Morris, 1969). The symmetry axis in the TI medium is made to coincide with the reference axis, and the crystal geometry factor W is obtained by expanding the distribution function in combination with the rock deposition process and stress. lmn The shale stiffness coefficient is obtained;

[0069]

[0070] in:

[0071]

[0072] Here C a It is obtained from the first step VRH model. During the synthesis process, the microstructure of shale changes, and the microstructure of shale is arranged in a horizontal direction. This is also a reduction of the deposition and formation process of shale in the geological history. The shale microstructure is a flaky horizontal structure. At this time, the shale exhibits vertical anisotropy (VTI, Vertical Transverse Isotropy).

[0073] Step 4: Substitute the shale model obtained in the third step into the Backus average model.

[0074] Backus (1962) showed that in the long wavelength limit, a layered medium composed of multiple layers of transversely isotropic materials (the symmetry axes of each layer are perpendicular to the layer) can be equivalent to a transversely isotropic medium, and its equivalent stiffness C* is:

[0075]

[0076] in:

[0077] A= <a-f 2 c -1 >+ <c -1 > -1 <fc -1 > 2

[0078] B= <b-f 2 c -1 >+ <c-1 > -1 <fc -1 > 2

[0079] C=<c -1 > -1

[0080] F=<c -1 > -1 <fc -1 >

[0081] D=<d -1 > -1

[0082] M= <m>

[0083] The brackets <.> indicate the average value of the closed properties weighted by volume proportion. a, b, c, d, f, and m are the elastic parameters of a single-layer transversely isotropic medium, and here they are substituted with the elastic parameters of the shale model obtained in the third step.

[0084] Step 5: Use the anisotropic parameters of the Thomsen model to obtain the longitudinal and transverse wave velocity models of anisotropic shale.

[0085] Thomsen (1986) proposed the following convenient notation for TI materials that are only weakly anisotropic. His notation uses the P- and S-wave velocities propagating along the symmetry axis from α and β, respectively, plus three other constants:

[0086]

[0087] Based on these constants, the three-phase speeds can be conveniently approximated as:

[0088]

[0089] V SH (θ)≈β(1+γsin 2 θ)

[0090] V P (θ)≈α(1+δsin 2 θcos 2 θ+εsin 4 θ)

[0091] Where θ is the angle of the wave vector relative to the x3 axis; V SH is the velocity of a pure transverse wave in the wavefront, in which there is no polarization component in the direction of the symmetry axis; V sv It is a pseudo-transverse wave in the direction of pure polarization; V p It is a pseudo longitudinal wave.

[0092] Through the above five steps, including: 1. Using the Voigt-Reuss-Hill average method (one of the synthesis methods), clay, quartz and calcite are synthesized into shale according to the elastic parameters and volume percentage of each mineral; 2. The geometric factors of the crystal orientation function are obtained using the porosity measured by logging; 3. The geometric factors of the crystal orientation function are brought into the matrix model to obtain the anisotropic matrix model of shale; 4. The shale model obtained in the third step is substituted into the Backus average model; 5. The longitudinal and transverse wave velocity models of the anisotropic shale are obtained using the anisotropic parameters of the Thomsen model, and the sample library data of a region can be obtained.

[0093] The present invention utilizes the elastic parameters and well logging data mentioned in Example 1;

[0094] Embodiment 2:

[0095] like Figure 1 As shown, the method of the present invention uses the geometric factor parameters to calculate the deep learning, and then displays the distribution value of the geometric factor in different stratum spaces in the target area to divide the under-compacted mudstone range and the well-compacted mudstone range, so as to identify the possible development range of low-speed mudstone, and initialize the corresponding deep learning network structure parameter value, and train the LSTM deep learning network through the training set and test set of the elastic parameters and logging data (porosity, permeability, water saturation, longitudinal and shear wave velocities as characteristic parameters, and geometric factors as target labels). In the training process, the gradient update algorithm of time truncated back propagation is used to iteratively calculate the correlation between the objective function and the geometric factor data set in the time dimension, complete the training of the LSTM deep learning network, and obtain the geometric factor prediction model.

[0096] In machine learning and deep learning models, target labels refer to the results that the model is trying to predict or learn during the training process. These labels are used to guide the model's learning process so that the model can capture the relationship between input data (features) and the corresponding output (prediction target). Specifically:

[0097] Define learning objectives:

[0098] The target label specifies what the model needs to predict during the training phase. In your scenario, the geometric factors are used as target labels, representing various mechanical properties and structural characteristics of mudstone. The model's task is to predict the values ​​of these factors based on the input features.

[0099] Train the model:

[0100] During the training process, the model compares the input features with the corresponding target labels, calculates the prediction error (loss function), and continuously adjusts the model parameters through the back-propagation algorithm to minimize the prediction error.

[0101] Target labels are critical for optimization during training, helping the model to continuously learn and improve its predictive capabilities.

[0102] Evaluate model performance:

[0103] After training is complete, the target labels are also used to evaluate the model’s predictive performance. By comparing the predictions output by the model on the test set with the actual target labels, a variety of evaluation metrics (such as root mean square error, mean absolute error, etc.) can be used to measure the effectiveness and accuracy of the model.

[0104] Guide decision making:

[0105] The trained model can be used to predict the geometric factors in new data sets in real time, thus providing a basis for geological exploration, mining decisions, etc.

[0106] Based on these predictions, researchers and engineers can develop more effective exploration and development plans.

[0107] In summary, the present invention proposes step 6: deep learning geometric factor prediction process based on multi-layer long short-term memory network:

[0108] The first layer is the forget gate, which is used to select and retain the previous step time C t-1 and the current step time C t The memory information of . The forget gate can be expressed as:

[0109] f t =σ(w f [s t-1 ,x t ]+b f )

[0110] In the formula, w f is the weight coefficient matrix of the forget gate layer, b f is the offset vector of the forget gate layer, w fs and w fx For input information t-1 and x t The corresponding weight coefficient matrix, f t is the forget gate at the current time t.

[0111] The second layer is the input gate structure, which controls the transmission of information to the cell state C and processes the input of the current sequence data. This layer consists of two parts. The first part uses the sigmoid function to determine the updated information state, while the second part uses the tanh unit to calculate a new vector:

[0112] i t =σ(w i [s t-1 ,x t ]+b i )

[0113]

[0114] where i t is the input gate, σ(·) is the sigmoid function, w i and b i are the weight coefficient matrix and offset vector of the input gate, respectively, t is the current cell state of the input layer.

[0115] The third layer is the cell state of unit C, which determines the addition and deletion of information through the forget gate and input gate. Multiply the cell state of the previous step by the forget gate f t , determine C t-1 The information that needs to be discarded is added to the information of the input gate to perform the final cell update:

[0116]

[0117] The last layer is the output gate structure. First, the sigmoid function is used to determine the cell unit output, then the tanh activation function is used to process the cell unit, and the sigmoid gate is used to calculate the output:

[0118] O t =σ(w o [s t-1 ,x t ]+b o )

[0119] S t =O t ⊙tanh(C t )

[0120] Where S t is the hidden state output, σ(·) is the sigmoid function, w o and b o is the weight matrix and the offset vector of the output gate.

[0121] The main training algorithm of LSTM is the back-propagation algorithm, which includes the following three steps:

[0122] 1) The output value of each neuron in the LSTM forward propagation is f t 、i t , C t , O t , S t .

[0123] 2) In LSTM, the propagation of the error term δ through reverse calculation includes the following directions: one is the back propagation over time; the other is the error term propagated to the previous layer.

[0124] 3) Calculate the gradient of each weight using the corresponding error term.

[0125] Furthermore, the principle is an algorithm for training recurrent neural networks (RNNs), which specifically expands the RNN only within a certain time step, calculates the gradient and updates the weight within these steps, without backpropagating the entire sequence, thereby reducing the computational complexity and alleviating the gradient vanishing problem. The objective function is a function of the difference between the model prediction and the actual target, the geometric factor data set is the change of the geometric factor in the time dimension with the change of the time step, and TBPTT is used in the time series to capture the temporal dependency between the geometric factors. First, the time series is divided into several time windows with a length of T. Backpropagation is performed in each window to reduce the burden of long sequence calculations. For each time window, the output of the objective function is calculated based on the current time step and the geometric factor. In each time window, the gradient of each time step is calculated by backpropagation. These gradients will help us adjust the parameters of the network, thereby gradually optimizing the objective function. After each time window, the network parameters are updated to minimize the objective function. At the same time, during the training process, the learning rate, batch sample size, weight reduction rate, and number of loop iterations of the deep learning network LSTM need to be set to achieve iterative optimization of the deep learning network. By comparing the changing trends of the training error and the test error of the LSTM deep learning network, the convergence consistency of the LSTM deep learning network training is judged, thereby determining the optimal deep learning network parameter values.

[0126] The geometric factor data is divided into a training data set and a test data set.

[0127] The training set is used to train the model, while the test set is used to detect the accuracy of the model and output the predicted geometric factors. The experimental data defines the original input parameters as X0 = {X1, X2, ..., X n }. Training set X train ={X1,X2,…,X r } and test set X test ={X1,X2,…,X e }Divided into 7:3.

[0128] The mean absolute error (MAE) and root mean square error (RMS error) were used to estimate the performance of the model.

[0129]

[0130] Among them, y i and p i are the actual geometric factor and the predicted geometric factor respectively, and N is the number of samples.

[0131] Embodiment 3:

[0132] In an alternative implementation: Data quality and preprocessing are critical steps when dealing with geoscience and geological research data, as these data often contain noise, incompleteness, and various inconsistencies. Therefore, a thorough preprocessing step can greatly improve the effectiveness of model training and the reliability of prediction results. Here are some specific steps and strategies that can be considered:

[0133] Data quality assessment

[0134] Data integrity check: Identify the location and number of missing values ​​in the data set. Understand whether the pattern of missing values ​​is random or if there is a trend, which may indicate a systematic problem in the data collection process.

[0135] Noise and outlier detection: Use statistical methods (e.g., z-score, IQR) to detect possible data outliers and noise.

[0136] Data consistency check: Ensure the consistency of units and scales of the same data variable in different data sources or datasets.

[0137] Data preprocessing steps

[0138] Handling missing values:

[0139] Deletion method: If the number of missing values ​​is small and random, you can consider deleting these records.

[0140] Filling method: Use the mean, median, most frequently occurring value (mode), or use interpolation methods (such as linear interpolation) to fill missing values.

[0141] Prediction method: Use simple models to predict missing values, such as regression, k-nearest neighbor filling, etc.

[0142] Outlier handling:

[0143] Confirmation: First, confirm whether the outliers are truly abnormal or are natural phenomena that exist in the data pattern.

[0144] Adjust or delete: Based on business and technical judgment, choose to keep, adjust (such as scaling), or delete outliers.

[0145] Data standardization and normalization:

[0146] Normalization: Scaling the data to the range [0,1] is suitable for situations where the input data feature range is very inconsistent.

[0147] Normalization: Adjusting the data to a normal distribution with mean 0 and standard deviation 1 is usually beneficial for algorithms that use gradient descent for optimization.

[0148] Data Dimensionality Reduction:

[0149] If the data dimension is too high, you can consider using techniques such as PCA (principal component analysis) to reduce the dimension, thereby reducing the computational complexity and the risk of overfitting.

[0150] Balance data:

[0151] If the class distribution in the dataset is uneven, you may need to balance the dataset using methods such as downsampling, oversampling, or synthesizing minority class instances (such as SMOTE).

[0152] Feature Selection and Extraction:

[0153] By calculating feature importance and performing correlation analysis, redundant and irrelevant features can be removed to improve model efficiency and effectiveness.

[0154] Time series data processing (if involved):

[0155] Ensure that the time steps of your time series data are consistent, interpolating or resampling if necessary.

[0156] Create time windows or features, such as moving averages, as needed to capture sequence characteristics.

[0157] quality assurance

[0158] Data Validation: After each preprocessing step, data validation is performed to ensure that the preprocessing process has not introduced new issues.

[0159] Documentation: Record each step of data preprocessing in detail to facilitate future review and reproduction.

[0160] By focusing on these data quality and preprocessing steps, you can ensure that the data fed into the model is high quality and potentially predictive, which can help improve model performance and prediction accuracy.

[0161] In an optional implementation: Hyperparameter tuning is an important step in optimizing machine learning and deep learning models. For LSTM models, its performance depends heavily on the choice of hyperparameters. Therefore, hyperparameter tuning of the model can significantly improve its predictive ability and generalization performance. The following are some commonly used hyperparameter tuning strategies:

[0162] Key Hyperparameters

[0163] Network structure parameters:

[0164] Number of LSTM layers: determines the depth of the network. Too few layers may not be able to capture complex features, while too many layers may lead to large computational complexity and overfitting.

[0165] Number of units per layer: The number of LSTM units or neurons in each layer. Generally, more units can capture more complex patterns, but it also increases the computational cost and the risk of overfitting.

[0166] Training parameters:

[0167] Learning rate: A hyperparameter that controls the step size of gradient descent. It usually needs to be carefully adjusted. If it is too large, it is easy to skip the optimal solution, while if it is too small, it will lead to slow convergence.

[0168] Batch size: affects the noise and speed of the training process. Too small will lead to unstable gradient updates, and too large will consume more memory.

[0169] Regularization parameter:

[0170] Dropout Rate: A means of preventing overfitting, achieved by randomly dropping units during training.

[0171] Other parameters:

[0172] Sequence step length: that is, the time step input to LSTM, which affects the length of context information seen by the model.

[0173] Activation function: such as tanh, ReLU, etc., affects the nonlinear expression ability of the model.

[0174] Hyperparameter Tuning Methods

[0175] Grid Search:

[0176] Define a discrete set of values ​​for hyperparameters and systematically try every possible combination. This is useful for a small range or a small number of parameters.

[0177] Random Search:

[0178] Randomly selecting parameters to try can often find good hyperparameter combinations faster than grid searching over a larger search space.

[0179] Bayesian Optimization:

[0180] The prior knowledge and model are used to update the probability distribution of the current search space, which can effectively handle continuous and discrete hyperparameters.

[0181] Hyperparameter Search Libraries:

[0182] Using specialized hyperparameter tuning tools, such as Optuna, Hyperopt, and Scikit-Optimize, can speed up and simplify the tuning process.

[0183] Cross Validation:

[0184] K-fold cross validation was used to evaluate the performance of each hyperparameter combination to reduce chance and better estimate the generalization ability of the model.

[0185] Practical advice

[0186] Stepwise tuning: Start tuning the main hyperparameters and gradually increase the complexity.

[0187] Computational cost considerations: Prioritize more efficient tuning strategies within the available computing resources.

[0188] Monitoring and recording: Recording each attempted hyperparameter combination and the corresponding evaluation results helps to understand and explain the final selected hyperparameters.

[0189] Early Stopping: Set the validation set to stop training early to avoid overfitting.

[0190] By properly selecting and tuning the hyperparameters of the LSTM model, the model's predictive performance and stability can be better enhanced. This not only helps improve the model's performance on the training set, but more importantly, enhances its generalization ability on unseen data.

[0191] In an optional implementation: Interpretability and validation are two very important aspects of machine learning models in practical applications, especially in fields such as geology and earth sciences, in order to ensure that the model results can be trusted and adopted. The following are detailed steps and strategies on how to improve model interpretability and validate results:

[0192] Explainability

[0193] Feature Importance Analysis:

[0194] Use methods such as feature importance plots or calculating Shapley values ​​to quantify the impact of input features on model predictions. This helps understand which features play a key role in the decision making process.

[0195] Visualization:

[0196] Use visualization tools such as t-SNE, PCA, etc. to show the distribution of high-dimensional data and how the model classifies or regresses it.

[0197] Drawing a visualization of the output of the model's intermediate layers helps us understand the model's feature abstraction process at each layer.

[0198] Sensitivity analysis:

[0199] By perturbing the input features and observing the changes in the prediction results, we can understand the sensitivity of the model to different features. This is a way to evaluate the robustness of the model.

[0200] Local Interpretability:

[0201] Use techniques such as LIME (Local Interpretable Model-agnostic Explanations) to provide interpretability for individual predictions and help understand local model behavior.

[0202] Model simplification:

[0203] Try using a more interpretable model like logistic regression or decision tree as a baseline for comparison, or simplify the model based on your understanding of the deep learning model.

[0204] verify

[0205] Cross Validation:

[0206] Use k-fold cross validation to evaluate the performance of the model to ensure that the model performs well on the entire dataset, not just on the training or test set.

[0207] Independent test set:

[0208] In addition to the regular training / testing split, keep a validation set independent of the parameter tuning and model development process to fairly evaluate the generalization ability of the model.

[0209] Time series validation (if applicable):

[0210] For time-related data, forward chaining is used to simulate the actual situation of continuously updating data.

[0211] Uncertainty Quantification:

[0212] When forecasting, estimate uncertainty (e.g. using Bayesian neural networks, Monte Carlo Dropout) and provide prediction intervals rather than just point forecasts.

[0213] Adversarial Verification:

[0214] Use adversarial examples or perturbed data to test the robustness and robustness of the model to ensure that the model does not respond drastically to small changes in the input.

[0215] Model diagnostics:

[0216] Diagnose the residual distribution of the model and check whether the fitting assumptions are met, such as normality and independence of the residuals.

[0217] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

[0218] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.< / m>

Claims

1. A method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters, characterized in that: The method comprises: Set up the LSTM multi-layer long short-term memory network structure, train it with the preset training set and test set, use the time truncated back propagation gradient update algorithm to iteratively calculate the correlation between the objective function and the geometric factor data set in the time dimension, complete the training of the LSTM network, and obtain the geometric factor prediction model; The mean absolute error (MAE) and root mean square error (RMS) were used to evaluate the model performance. Finally, the optimized LSTM network was used to predict the geometric factors and determine the distribution range of undercompacted mudstone and its compaction degree.

2. The method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters according to claim 1, characterized in that: The LSTM multi-layer long short-term memory network structure is set up, specifically including: The first layer is the forget gate, which is used to select and retain the previous step time C t-1 and the current step time C t The memory information of the forget gate is expressed as: f t =σ(w f [s t-1 ,x t ]+b f ) In the formula, w f is the weight coefficient matrix of the forget gate layer, b f is the offset vector of the forget gate layer, w fs and w fx For input information t-1 and x t The corresponding weight coefficient matrix, f t is the forget gate at the current time t; The second layer is the input gate structure; the input gate controls the transmission of information to the cell state C and processes the input of the current sequence data; this layer consists of two parts, the first part uses the sigmoid function to determine the updated information state, and the second part uses the tanh unit to calculate a new vector: I t =σ(w i [s t-1 ,x t ]+b i ) where i t is the input gate, σ(·) is the sigmoid function, w i and b i are the weight coefficient matrix and offset vector of the input gate, respectively, t is the current cell state of the input layer; The third layer is the cell state of unit C; the addition and deletion of information are determined by the forget gate and input gate; the cell state of the previous step is multiplied by the forget gate f t , determine C t-1 The information that needs to be discarded is added to the information of the input gate to perform the final cell update: The last layer is the output gate structure. The sigmoid function is used to determine the cell unit output, the tanh activation function is used to process the cell unit, and the sigmoid gate is used to calculate the output: The t =σ(w o [s t-1 ,x t ]+b o ) S t =O t ⊙tanh(C t ) Where S t is the hidden state output, σ(·) is the sigmoid function, w o and b o is the weight matrix and the offset vector of the output gate.

3. The method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters according to claim 1, characterized in that: The elastic parameters and logging data are organized into training sets and test sets. The data used as input features include porosity, permeability, water saturation, P-wave velocity, and S-wave velocity. The data used as target labels are geometric factors obtained through modeling. The data are divided into training data sets and test data sets to evaluate model performance.

4. The method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters according to claim 1, characterized in that: The gradient update algorithm of time-truncated back-propagation is used to iteratively calculate the relationship between the objective function and the geometric factor data set in the time dimension, including: The output value of each neuron in the LSTM forward propagation is f t 、i t , C t , O t , S t ; In LSTM, the propagation of the error term δ through reverse calculation includes the following directions: one is the back propagation over time; the other is the error term propagated to the previous layer; the gradient of each weight is calculated using the corresponding error term.

5. The method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters according to claim 1, characterized in that: The mean absolute error (MAE) and root mean square error (RMS) are used to evaluate the model performance, including: Define the original input parameters as X0 = {X1, X2, ..., X n }, training set X train ={X1,X2,…,X r } and test set X test ={X1,X2,…,X e }Divided into 7:3; The mean absolute error (MAE) and root mean square error (RMS error) are used to estimate the performance of the model: Among them, y i and p i are the actual geometric factor and the predicted geometric factor respectively, and N is the number of samples.

6. The method for predicting the distribution of undercompacted low-velocity mudstone based on geometric factor parameters according to claim 1, characterized in that: After training, the LSTM network can predict the target area through the learned model relationship and predict the corresponding geometric factors; combined with geological characteristics and sedimentary environment, it can identify the distribution and maturity of undercompacted low-speed mudstone.

Images