Geophysical method for predicting coal rock thickness

Abstract

The invention discloses a geophysical method for predicting coal rock thickness, particularly relates to the technical field of coalbed methane exploration, and has the core innovation that a wedge-shaped geologic model containing nine kinds of roof and floor lithology combinations is constructed, and seismic attributes sensitive to a thin coal seam are optimized through principal component analysis; converting the complex relationship between the seismic attributes and the coal thickness into a linear separable problem by using the specific high-dimensional nonlinear mapping capability of the radial basis function neural network; a three-level closed-loop mechanism of attribute optimization-network training-dynamic verification is established, and model self-optimization is realized through spider diagram analysis, over-fitting monitoring and new well triggering iteration. According to the method, the tuning distortion problem of a traditional linear model in thin coal seam prediction is solved, the reliability of well-free area prediction is remarkably improved, and a key technical support is provided for coal seam gas dessert identification and development decision making.

Description

Technical Field

[0001] This invention relates to the field of coalbed methane exploration technology, and more specifically, to a geophysical method for predicting coal and rock thickness. Background Technology

[0002] Coalbed methane (CBM), as a highly efficient and clean energy source, is abundant in my country and its exploration and development are receiving increasing attention. CBM production is directly related to coal seam thickness, and accurate prediction of coal thickness is crucial for reserve assessment, well location deployment, and economic benefits. Traditional prediction methods have evolved from early reliance on borehole interpolation to seismic reflection wave time-thickness conversion, and then to using seismic amplitude to predict coal thickness. While these methods have achieved some success, they are limited by the sparsity of well logging data and the generally thin geological characteristics of coal seams (mostly <1.5 meters). Existing methods exhibit significant prediction errors in areas without wells, making it difficult to meet the stringent vertical resolution requirements of CBM exploration and development.

[0003] Current mainstream technologies suffer from three major bottlenecks: First, borehole interpolation-based methods have low reliability in well-free areas, especially lacking sensitivity to spatial variations in thin coal seams. Second, seismic amplitude prediction methods are affected by thin-layer tuning effects; when the coal seam thickness is less than λ / 4 (λ is the seismic wavelength), the amplitude-thickness relationship exhibits nonlinear distortion, leading to a sharp drop in prediction accuracy. Third, traditional linear statistical models (such as multiple regression) cannot effectively characterize the complex nonlinear mapping relationship between coal seam thickness and seismic attributes, and the attribute optimization process lacks systematic analysis of changes in the lithological combination of the roof and floor, limiting the model's generalization ability. These shortcomings result in existing methods having a mean absolute error generally greater than 1.2 meters, severely restricting the identification of sweet spots in coalbed methane.

[0004] Therefore, a geophysical method for predicting coal and rock thickness is proposed to address the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a geophysical method for predicting coal and rock thickness to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a coal seam thickness prediction method based on radial basis function neural networks, comprising: jointly analyzing well logging data and 3D seismic data of the work area, establishing a wedge-shaped geological model containing multiple roof and floor lithological combinations based on seismic forward modeling, and identifying seismic attribute parameters that are significantly correlated with changes in coal seam thickness; selecting an adaptive time window along the roof and floor interface of the target coal seam based on refined stratigraphic interpretation, systematically extracting amplitude, waveform, phase, and spectral statistical seismic attribute volumes, and establishing a quantitative relationship model between seismic attribute values ​​at the drilling location and the measured coal seam thickness by combining high-precision well-seismic calibration technology; implementing a seismic attribute optimization process, firstly eliminating weakly correlated attributes with correlation coefficients lower than a set threshold through single-well attribute and coal thickness correlation analysis, and secondly calculating the cross-correlation matrix between attributes and eliminating redundant attributes with cross-correlation coefficients greater than a preset value. Finally, principal component analysis was used to reduce the dimensionality of the remaining attributes, and the sensitive attribute combination with the highest principal component loading weight was selected to generate a spider diagram for sensitivity verification, forming the final sensitive attribute set. The selected sensitive seismic attribute data and the corresponding well point coal seam thickness samples were jointly input into the radial basis function neural network for end-to-end supervised training. This network maps the input samples to a high-dimensional space through hidden layer radial basis functions to establish a nonlinear mapping relationship between seismic attributes and coal seam thickness. The prediction error was verified using measured coal thickness data from independent test wells. When the mean absolute error exceeded the industry allowable threshold, the attribute optimization step was returned to re-select sensitive attributes. Based on the trained neural network model, the optimized sensitive attribute set was extracted from the seismic data volume of the entire work area. After data normalization preprocessing, it was input into the network to predict the planar distribution of coal seam thickness and generate error analysis maps for geological interpretation.

[0007] Preferably, the specific implementation of the principal component analysis method includes: constructing an original data matrix with coal seam thickness variation as rows and seismic attribute parameters as columns; standardizing the attribute values ​​in each column of the matrix to eliminate dimensional differences; calculating the covariance characteristics of the standardized matrix and decomposing it to obtain eigenvalues ​​and corresponding eigenvectors; selecting principal components with a cumulative contribution rate exceeding 85% by sorting the eigenvalues ​​from largest to smallest; identifying the seismic attributes that have the most significant impact on the changes of principal components by analyzing the absolute values ​​of the eigenvector components; the screening process must simultaneously meet the dual conditions that the principal component load weight is greater than 0.7 and the correlation coefficient with coal seam thickness is greater than 0.5.

[0008] Preferably, the training process of the radial basis function neural network includes: performing range normalization on the input multidimensional sensitive seismic attribute data to eliminate the order-of-magnitude differences between different attributes; adaptively determining the hidden layer basis function centers using a nearest neighbor clustering learning algorithm, and iteratively updating the center points by dynamically calculating the Euclidean distance between new samples and existing cluster centers. When the minimum distance is less than the Gaussian function width threshold, the sample is assigned to the nearest neighbor category and the category output mean is updated; otherwise, a new cluster center is created; adaptively optimizing the basis function width parameter based on the prediction error function value, and re-clustering by iteratively reducing the width threshold until the error converges to a set range; the output layer adopts a linear weighted summation calculation method, and its weight value is the set of the output means of each category of samples, ultimately forming a continuous mapping function from the seismic attribute space to the coal seam thickness.

[0009] Preferably, the dynamic update mechanism of the nearest neighbor clustering learning algorithm is specifically manifested as follows: the first training sample is used as the initial cluster center to establish the first hidden layer node. For each subsequent new sample, the minimum Euclidean distance from it to all existing cluster centers is calculated. If the distance is less than the current Gaussian function width threshold, the sample is assigned to the corresponding category. At the same time, the cumulative value of the output vector and the sample counter of the category are updated and the weights are recalculated. If the minimum distance exceeds the width threshold, the sample is used as the new cluster center to create a hidden layer node. The total number of hidden layer nodes in the network increases adaptively with the complexity of the sample space distribution. The output weight of each node is always the arithmetic mean of the measured coal seam thickness of all samples in the category.

[0010] Preferably, the basis function width optimization adopts an error feedback control strategy: the Gaussian function width is initially set to 1 / 2 of the maximum distance in the training sample space, the sum of squared prediction errors of all training samples under the current parameters is calculated, and when the error value is greater than the convergence threshold, the width parameter is reduced according to the exponential decay rule and clustering learning is re-executed. The iterative process continues until the error function value is less than the set threshold or the maximum number of iterations is reached. The final width parameter enables the network to achieve the optimal balance between prediction accuracy and computational efficiency.

[0011] Preferably, the seismic forward modeling employs a parametric wedge geological model. This model systematically characterizes the continuous variation of coal seam thickness from 0.5 meters to 10 meters and covers nine typical roof and floor lithological combinations: both roof and floor are mudstone, both are limestone, both are sandstone, roof is mudstone and floor is limestone, roof is mudstone and floor is sandstone, roof is limestone and floor is mudstone, roof is limestone and floor is sandstone, roof is sandstone and floor is mudstone, and roof is sandstone and floor is limestone. By analyzing the response sensitivity of seismic attributes to coal seam thickness under different lithological combinations through forward modeling and synthesizing seismic records, a theoretical basis is provided for the optimal selection of attributes in actual work areas.

[0012] Preferably, the verification mechanism runs through the entire process: in the attribute optimization stage, the discriminative power of sensitive attribute combinations is analyzed through spider diagram visualization; in the neural network training stage, 20% of well points are reserved as test wells to monitor overfitting, requiring that the average relative error between the training set and the test set is less than 15%; in the work area prediction stage, cross-validation is used, and when the absolute error between the newly added drilling data and the prediction result exceeds 0.8 meters, the model optimization process is automatically triggered to ensure that the prediction result meets the accuracy requirements for coalbed methane exploration and development.

[0013] The technical effects and advantages of this invention are as follows:

[0014] Compared to existing technologies, this invention utilizes the unique high-dimensional spatial nonlinear mapping capability of radial basis function neural networks to transform the complex relationship between seismic attributes and coal seam thickness into a linearly separable problem, overcoming the limitations of traditional linear models in predicting thin coal seams. Based on a wedge-shaped geological model-based seismic attribute optimization mechanism, the invention systematically quantifies the modulation effect of different roof and floor lithological combinations on seismic response, significantly reducing the impact of thin-layer tuning distortion on prediction results. Combined with a three-level closed-loop workflow of "attribute optimization - network training - dynamic verification," the invention achieves self-correction and continuous optimization of coal seam thickness prediction in well-free areas, significantly improving the generalization ability of geological models while ensuring vertical resolution, and providing highly reliable technical support for coalbed methane sweet spot identification and horizontal well trajectory design. Attached Figure Description

[0015] Figure 1 This is a flowchart of the coal seam thickness prediction method of the present invention.

[0016] Figure 2 This is a diagram of the radial basis neural network topology of the present invention.

[0017] Figure 3 This is a schematic diagram of the three-level verification mechanism of the present invention.

[0018] Figure 4 A flowchart of the intelligent coal seam thickness prediction method provided in the embodiments of the present invention;

[0019] Figure 5 Spider diagrams illustrating the statistical sensitivity of coal thicknesses with different seismic attributes, provided in embodiments of the present invention;

[0020] Figure 6 A topology diagram of a radial basis function neural network provided in an embodiment of the present invention;

[0021] Figure 7 The diagram shows the error analysis of predicting coal seam thickness using a radial basis function neural network, as provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1

[0024] As attached Figure 1-3 The method for predicting coal seam thickness based on radial basis function neural networks includes: jointly analyzing well logging data and 3D seismic data in the work area; establishing a wedge-shaped geological model containing various roof and floor lithological combinations based on seismic forward modeling; identifying seismic attribute parameters that are significantly correlated with changes in coal seam thickness; selecting an adaptive time window along the roof and floor interface of the target coal seam based on refined stratigraphic interpretation; systematically extracting amplitude, waveform, phase, and spectral statistical seismic attribute volumes; and establishing a quantitative relationship model between seismic attribute values ​​at the drilling location and the measured coal seam thickness using high-precision well-seismic calibration technology; implementing a seismic attribute optimization process, firstly eliminating weakly correlated attributes with correlation coefficients below a set threshold through single-well attribute and coal thickness correlation analysis; secondly calculating the cross-correlation matrix between attributes and eliminating redundant attributes with cross-correlation coefficients greater than a preset value; and finally using principal component analysis. The analytical method performs feature dimensionality reduction on the remaining attributes, selects the sensitive attribute combination with the highest principal component loading weight to generate a spider diagram for sensitivity verification, and forms the final sensitive attribute set. The selected sensitive seismic attribute data and the corresponding well point coal seam thickness samples are jointly input into the radial basis function neural network for end-to-end supervised training. This network maps the input samples to a high-dimensional space through hidden layer radial basis functions to establish a nonlinear mapping relationship between seismic attributes and coal seam thickness. The prediction error is verified using measured coal thickness data from independent test wells. When the mean absolute error exceeds the industry allowable threshold, the attribute optimization step is returned to re-select sensitive attributes. Based on the trained neural network model, the optimized sensitive attribute set is extracted from the seismic data volume of the entire work area. After data normalization preprocessing, it is input into the network to predict the planar distribution of coal seam thickness and generate error analysis maps for geological interpretation.

[0025] In implementing this method in coalbed methane exploration areas, firstly, 3D seismic data and well logging data are collected. A wedge-shaped geological model (e.g., mudstone-coal seam-limestone) containing nine roof and floor lithological combinations is established through seismic forward modeling, identifying seismic attributes with a correlation higher than 0.6 with coal seam thickness. Dynamic time windows are selected along the target stratigraphic level (calculation formula: time window length = 2 × coal seam time thickness + 10ms), extracting 42 attributes such as amplitude and waveform. After three optimization steps: 1) attributes with a correlation coefficient <0.55 with coal thickness are removed; 2) redundant attributes with a cross-correlation coefficient >0.8 are deleted; 3) principal component analysis is used to reduce the dimensionality to sensitive attribute combinations with a cumulative contribution rate >85% (e.g., spectral attenuation, waveform curvature). The optimized attributes and the measured coal thickness of 23 wells are input into the radial basis function neural network. The nearest neighbor clustering algorithm is used to adaptively generate the hidden layer center. When the average absolute error of the test well prediction is >0.8 meters, the optimized attributes are returned. After the error is qualified, the coal seam thickness distribution is predicted for the seismic data of the entire work area, and a thickness planar map with a grid accuracy of 25m×25m is output for well location deployment.

[0026] In this embodiment, the specific implementation of the principal component analysis method includes: constructing an original data matrix with coal seam thickness variation as rows and seismic attribute parameters as columns; standardizing the attribute values ​​in each column of the matrix to eliminate dimensional differences; calculating the covariance characteristics of the standardized matrix and decomposing it to obtain eigenvalues ​​and corresponding eigenvectors; selecting principal components with a cumulative contribution rate exceeding 85% by sorting the eigenvalues ​​from largest to smallest; and identifying the seismic attributes that have the most significant impact on the changes of principal components by analyzing the absolute values ​​of the eigenvector components. The screening process must simultaneously meet the dual conditions that the principal component load weight is greater than 0.7 and the correlation coefficient with coal seam thickness is greater than 0.5.

[0027] Seismic attribute optimization specifically includes: constructing a matrix with 20 coal seam thickness points as rows and 32 seismic attributes as columns; standardizing the matrix to eliminate dimensional differences using the formula Z = (X - μ) / σ; calculating the covariance matrix and decomposing the eigenvalues; selecting principal components with eigenvalues ​​>1.2 and cumulative contribution rates >85%; and screening attributes with absolute values ​​of eigenvector components >0.7 (e.g., dominant frequency load 0.82, waveform complexity 0.79). Sensitivity analysis using spider diagrams ensures that the attribute combinations have a discrimination effect >75% on thin coal seams (<2m), and finally retaining 7 core attributes in the attribute set.

[0028] In this embodiment, the training process of the radial basis function neural network includes: performing range normalization on the input multidimensional sensitive seismic attribute data to eliminate the order-of-magnitude differences between different attributes; adaptively determining the hidden layer basis function centers using a nearest neighbor clustering learning algorithm, and iteratively updating the center points by dynamically calculating the Euclidean distance between new samples and existing cluster centers. When the minimum distance is less than the Gaussian function width threshold, the sample is assigned to the nearest neighbor category and the category output mean is updated; otherwise, a new cluster center is created; adaptively optimizing the basis function width parameter based on the prediction error function value, and re-clustering by iteratively reducing the width threshold until the error converges to a set range; the output layer adopts a linear weighted summation calculation method, and its weight value is the set of the output mean values ​​of each category of samples, ultimately forming a continuous mapping function from the seismic attribute space to the coal seam thickness.

[0029] The training process for the radial basis function neural network is as follows: 1) Input data normalization: X_norm=(X-X_min) / (X_max-X_min), amplitude classes are normalized to [0,1], and frequency classes are normalized to [-1,1]; 2) Nearest neighbor clustering initialization: the first sample is used as the first center. When a new sample satisfies min||Xk-ci||≤0.4*r, the weights w_i=A(l) / B(l) of the class center are updated; otherwise, a new center is added; 3) Basis function width optimization: initially r=maximum distance in the sample space / 2, according to the error function E=Σ(y_pred-y_true). 2 Iterative adjustment: when E>0.15, let r=0.85*r to re-cluster until convergence; 4) Linear weighted calculation of coal thickness prediction value of output layer.

[0030] In this embodiment, the dynamic update mechanism of the nearest neighbor clustering learning algorithm is specifically manifested as follows: the first training sample is used as the initial cluster center to establish the first hidden layer node. For each subsequent new sample, the minimum Euclidean distance from it to all existing cluster centers is calculated. If the distance is less than the current Gaussian function width threshold, the sample is assigned to the corresponding category. At the same time, the cumulative value of the output vector and the sample counter of the category are updated and the weights are recalculated. If the minimum distance exceeds the width threshold, the sample is used as the new cluster center to create a hidden layer node. The total number of hidden layer nodes in the network increases adaptively with the complexity of the sample space distribution. The output weight of each node is always the arithmetic mean of the measured coal seam thickness of all samples in the category.

[0031] The nearest neighbor clustering algorithm is implemented as follows: Using the sample data from well X1 as the initial cluster center c1, the Euclidean distance between each subsequent sample Xk and all centers is calculated, and the minimum distance d_min is found. If d_min ≤ 0.45*r (where r is the current Gaussian function width), Xk is assigned to the corresponding category, and the cumulative output vector value A(l) and sample counter B(l) of that category are updated, with weight w_i = A(l) / B(l). If d_min > 0.45*r, a hidden layer node is created with Xk as the new center c_new. Finally, the weight of each node is the arithmetic mean of the measured coal thickness values ​​of all samples in that category.

[0032] In this embodiment, the basis function width optimization adopts an error feedback control strategy: the initial Gaussian function width is set to 1 / 2 of the maximum distance in the training sample space, the sum of squared prediction errors of all training samples under the current parameters is calculated, and when the error value is greater than the convergence threshold, the width parameter is reduced according to the exponential decay rule and clustering learning is re-executed. The iterative process continues until the error function value is less than the set threshold or the maximum number of iterations is reached. The final width parameter enables the network to achieve the optimal balance between prediction accuracy and computational efficiency.

[0033] The basis function width optimization employs an error feedback mechanism: The initial width r = 5.2 (half of the maximum distance in the sample space, 10.4) is calculated, and the sum of squared prediction errors E for all training samples in the current cluster is calculated. When E > 0.15, r = 0.85*r is updated according to the exponential decay rule, and clustering learning is re-executed, recording the error change rate in each iteration. If the error decrease rate is < 5% or E ≤ 0.12 for three consecutive iterations, the process stops, ultimately obtaining the optimal r = 2.1, achieving a balance between prediction accuracy and computational efficiency.

[0034] In this embodiment, the seismic forward modeling uses a parametric wedge geological model. This model systematically characterizes the continuous variation of coal seam thickness from 0.5 meters to 10 meters and covers nine typical roof and floor lithological combinations: both roof and floor are mudstone, both are limestone, both are sandstone, roof is mudstone and floor is limestone, roof is mudstone and floor is sandstone, roof is limestone and floor is mudstone, roof is limestone and floor is sandstone, roof is sandstone and floor is mudstone, and roof is sandstone and floor is limestone. By analyzing the response sensitivity of seismic attributes to coal seam thickness under different lithological combinations through forward modeling and synthesizing seismic records, a theoretical basis is provided for the optimal selection of attributes in actual work areas.

[0035] Seismic forward modeling employed a parametric wedge model: coal seam thickness gradually varied from 0.5m to 10.0m in 0.5m intervals, with nine roof and floor lithological combinations configured (including mudstone for both roof and floor, limestone for roof, and sandstone for floor, etc.). Model parameters were set as follows: coal seam velocity 2200m / s, mudstone velocity 2800m / s, limestone velocity 4200m / s, and sandstone velocity 3500m / s. Seismic records were synthesized using 25Hz Ricker wavelet with a signal-to-noise ratio controlled at 3:1 to analyze the response characteristics of properties to coal thickness under different lithological combinations.

[0036] In this embodiment, the verification mechanism runs through the entire process: during the attribute optimization stage, the discriminative power of sensitive attribute combinations is analyzed through spider diagram visualization; during the neural network training stage, 20% of well points are reserved as test wells to monitor overfitting, requiring that the average relative error between the training set and the test set is less than 15%; during the work area prediction stage, cross-validation is used, and when the absolute error between the newly added drilling data and the prediction result exceeds 0.8 meters, the model optimization process is automatically triggered to ensure that the prediction result meets the accuracy requirements for coalbed methane exploration and development.

[0037] The three-level validation mechanism is implemented as follows: 1) In the attribute optimization stage, the discrimination index is determined to be met when the radius ratio of sensitive attribute combinations shown in the spider diagram is >1.5; 2) In the neural network training stage, 20% of well points are reserved as a test set, requiring the average relative error of the training set to be <12% and the test set <15%, and overfitting is monitored; 3) In the work area prediction stage, model optimization is automatically triggered when the absolute error between newly added drilling data and prediction results is >0.8 meters. If two consecutive wells exceed the error, the network is retrained. The final model must meet all three-level validation standards before it can be applied to exploration and development.

[0038] Example 2

[0039] The present invention will now be described in detail with reference to the accompanying drawings and examples. Figure 4-7 As shown, it includes the following steps:

[0040] Step 1: Analyze well logging and seismic data to establish a theoretical geological model. Ideally, information such as the depositional environment, lithology, and thickness of coal seams, as well as seismic wavelength attribute parameters, can be directly or indirectly reflected using seismic attributes. However, due to various uncontrollable interferences such as noise, each attribute has different sensitivities to predicting coal seam physical properties. Therefore, it is first necessary to identify attributes with a high correlation to coal seam thickness using forward modeling of seismic attributes. A model was established as follows... Figure 1The theoretical geological model shown, the wedge model, reflects the variation law of coal seam thickness from thin to thick. Adding the lithological variations of the coal seam's roof and floor and model parameters suitable for the work area, theoretical geological models under different geological conditions can be obtained. These mainly include the following nine types of roof and floor lithological contact relationships: ① Coal seam roof and floor are mudstone; ② Mudstone-coal seam-limestone contact relationship; ③ Mudstone-coal seam-sandstone contact relationship; ④ Limestone-coal seam-mudstone contact relationship; ⑤ Limestone-coal seam-limestone contact relationship; ⑥ Limestone-coal seam-sandstone contact relationship; ⑦ Sandstone-coal seam-sandstone contact relationship; ⑧ Sandstone-coal seam-mudstone contact relationship; ⑨ Sandstone-coal seam-limestone contact relationship.

[0041] Step 2: Select an appropriate time window to extract various seismic attributes along the target stratum, including amplitude, waveform, phase, and spectral statistics, and analyze the relationship between seismic attributes at the drilling location and coal seam thickness. The main factors influencing the accuracy of seismic attributes in reflecting geological conditions are the interpretation accuracy of the target stratum and the selection of the time window. Therefore, it is necessary to conduct detailed well-seismic calibration and stratum interpretation in the high-resolution processed seismic data to avoid prediction errors caused by inaccurate extracted attribute values.

[0042] Step 3: Seismic attribute optimization. Attribute optimization involves selecting attributes that are more sensitive to coal thickness from dozens of options for attribute analysis. Having more seismic attribute parameters does not necessarily improve coal thickness prediction. Seismic attribute optimization can reduce attribute dimensionality, obtain the optimal attribute combination, improve accuracy, and save time. Methods mainly include single-well and attribute correlation analysis, cross-correlation analysis between different attributes, and principal component analysis for data dimensionality reduction. By combining the coal thickness distribution at the monitoring well location, through correlation statistics and principal component analysis to reduce the feature dimension, attribute combinations that are more sensitive to coal thickness are selected from all attributes and statistically analyzed into a spider diagram. These sensitive attributes are then stored in a single attribute set.

[0043] To ensure the relative independence of each attribute and the stability of the algorithm, cross-correlation analysis of seismic attributes is performed. Based on the correlation coefficients between each seismic attribute, and referring to the correlation coefficient between seismic attributes and coal thickness, attributes with large correlation coefficients between seismic attributes and those with small correlation coefficients between seismic attributes and coal thickness are eliminated.

[0044] Because seismic attributes exhibit parametric ambiguity, a single attribute, whether qualitative or quantitative, may not be accurate. Principal Component Analysis (PCA) is a classic unsupervised linear feature extraction algorithm that evaluates the sensitivity of each seismic attribute to variations in coal seam thickness, thereby selecting the most sensitive attributes that can distinguish between different overlay relationships.

[0045] The steps for implementing PCA are as follows:

[0046] 1. Standardized data

[0047]

[0048] Where xij represents the j-th seismic attribute under the i-th coal seam thickness, n is the number of changes in coal thickness, and p is the number of seismic attributes.

[0049]

[0050] Where μ j It is the mean of the j-th seismic attribute, σ j It is its standard deviation.

[0051] 2. Perform principal component analysis on the standardized matrix Z and calculate its covariance matrix C:

[0052]

[0053] The covariance matrix C is decomposed into eigenvalues ​​and eigenvectors. The eigenvectors correspond to the directions of the principal components, while the eigenvalues ​​characterize the variance of each principal component. By examining the principal component loadings (the components of the eigenvectors), we can determine which seismic attributes contribute most to the principal components, thus selecting the most sensitive attributes.

[0054] Step 4: Input the multiple sensitive seismic attributes selected in Step 3 and the statistical wellpoint coal seam thickness samples into the radial basis function neural network. The network transforms the input samples from low dimension to high dimension according to the characteristics of seismic attributes and wellpoint coal thickness, so that features that are linearly inseparable in low-dimensional space become linearly separable in high-dimensional space. Output the prediction result and check whether it meets the error range by combining it with the coal seam thickness of the test well. If it exceeds the error allowable range, it means that the selected sensitive attributes are not suitable and need to be further optimized; if it is within the error allowable range, it means that the sensitive attributes selected in Step 3 are sufficient to distinguish the predicted coal seam thickness distribution and can be used in actual work areas.

[0055] Step 5: Extract the sensitive seismic attributes selected in Step 3 and Step 4 from the actual seismic data along a suitable time window, preprocess them, and input them into the radial basis function neural network to obtain the prediction results. Combine the well logging data of the test well to determine whether the coal seam thickness distribution predicted by the radial basis function neural network meets the conditions.

[0056] The radial basis function (RBF) neural network mentioned in steps 3 and 4 is a supervised feedforward neural network that uses radial basis functions as basis functions for hidden layer neurons. It features nonlinear optimal approximation, and it has been proven that an RBF network with a sufficient number of hidden layer neurons can approximate any continuous function with arbitrary precision. Its neural network topology is as follows: Figure 2 As shown, it includes an input layer, a hidden layer, and an output layer.

[0057] The input layer receives input samples and passes them to the hidden layers. The hidden layers are a set of radially symmetric kernel functions, which form a "basis" for the input samples, transforming them from low-dimensional to high-dimensional space, making features that are linearly inseparable in low-dimensional space linearly separable in high-dimensional space. The output layer is a linear function that outputs the prediction result. The steps of a radial basis function neural network are as follows:

[0058] 1. Preprocess the input sensitive seismic attributes. Different types of seismic attributes often have different value ranges. The purpose of normalization is to standardize the input seismic attributes to ensure that the numerical ranges of each attribute are similar, thereby avoiding any attribute from having an excessive impact on the results. The formula is as follows:

[0059]

[0060] Where A represents the seismic attribute value after range preprocessing; A f Represents the original seismic attribute data; A max A represents the maximum value of the original seismic attribute data; min This represents the minimum value of the original seismic attribute data.

[0061] 2. Activate the hidden layer radial basis functions. The connection from the input layer to the hidden layer is a fixed connection with weights of 1. The hidden layer consists of a set of radial basis functions, the most commonly used being the Gaussian function, i.e.

[0062]

[0063] In the formula, x k c represents the training sample; i The center of the basis function is represented by σ; the width of the basis function around the center point is represented by σ; k = 1, 2, 3, ..., n represents the number of training samples formed by seismic attribute values ​​at the well point; i = 1, 2, ..., m represents the number of basis function centers; ||x k -c i || represents the norm of a vector, which is x k With c i The distance between them; f(x) k ) in c i There is a unique maximum value at ||x k -c i As || increases, f(x) k It decays rapidly to zero. For a given input x k ∈R n Only a small portion are close to x k The center is activated.

[0064] 3. The nearest neighbor clustering learning algorithm adaptively selects the basis function centers. The training samples and output results are combined to form a data pair, defined as (x...k ,y k ), where k = 1, 2, 3, ... n; Define a vector A(l) to store the sum of the output vectors of each category, and define a counter B(l) to count the number of samples belonging to each category, where l is the number of categories; Set an appropriate Gaussian function width r, then, starting from the first data pair (x1, y1), establish a cluster center on x1, let c1 = x1, A(1) = y1, B(1) = 1. At this time, there is only one hidden layer node in the RBF network, the cluster center of this node is c1, and the weight value between this hidden layer node and the output layer is w1 = A(1) / B(1).

[0065] For the second sample data pair (x2, y2), calculate the distance |x2-x1| from x2 to the cluster center c1. If |x2-x1|≤r, then c1 is the nearest neighbor cluster center of x2. In this case, let A(1)=y1+y2, B(1)=2, w1=A(1) / B(1); if |x2-c1|>r, then x2 is taken as a new cluster center, let c2=x2, A(2)=2, B(2)=1, and add a new hidden layer node in the RBF network. The weight of this hidden layer node and the output layer is w2=A(2) / B(2). And so on, considering up to the kth sample data pair (x2, y2), we can calculate the distance |x2-x1| from x2 to the cluster center c1. k ,y k When x is calculated, find x respectively. k To cluster center c i Distance |x k -c i |, where i = 1, 2, ..., m represents the index of the existing cluster centers in the RBF network. Let |x k -c j | is the minimum distance among them, if |x k -c j |≤r, i.e., c j For x k Let A(m) = A(m) + y be the nearest neighbor cluster center, j∈(i|i=1,2,...,m)…. k B(m) = B(m) + y k w m =A(m) / B(m); if |x k -c j |>r, then x k As a new cluster center, let c m+1 =x k A(m+1)=y k B(m+1) = 1, and add the (m+1)th hidden layer node to the RBF network. The weight of this hidden layer node and the output layer is w. m+1= A(m+1) / B(m+1). In this way, the center of the basis function can be adaptively selected, that is,

[0066]

[0067] Step four: Adaptively determine the basis function width using the error function. The size of the Gaussian function width *r* determines the complexity of the RBF network; that is, the smaller *r* is, the more clusters the network obtains, and the greater the computational cost. The optimal Gaussian function width *r* can be determined based on the error function, i.e.,

[0068]

[0069] In the formula, k = 1, 2, 3, ..., n represents the index of the training samples composed of seismic attribute values ​​at the well point; α represents the number of iterations. If the error function E α If the error function E is greater than the preset convergence threshold ε, then decrease the width r of the Gaussian function and return to step three; if the error function E α When the value is less than the threshold ε, the learning process ends. At this point, the network adaptively selects the optimal basis function width r.

[0070] Step 5: Output the prediction result. Based on the hidden layer basis function centers learned from the training samples and their weights with the output layer, a linear transformation is performed to output the prediction result, i.e.

[0071]

[0072] In the formula, w i The weights of the hidden layer basis function centers and the output layer are represented by k = 1, 2, 3, ..., n, which represent the number of training samples formed by the seismic attribute values ​​at the well point; and i = 1, 2, ..., m, which represent the number of hidden layer basis function centers.

[0073] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0074] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0075] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting coal seam thickness based on radial basis function neural networks, characterized in that, include: By jointly analyzing well logging data and 3D seismic data from the work area, a wedge-shaped geological model containing various roof and floor lithological combinations was established based on seismic forward modeling, identifying seismic attribute parameters that are significantly correlated with coal seam thickness variations. Based on refined stratigraphic interpretation, an adaptive time window was selected along the roof and floor interfaces of the target coal seam, and the system extracted amplitude, waveform, phase, and spectral statistical seismic attribute volumes. Combined with high-precision well-seismic calibration technology, a quantitative relationship model between seismic attribute values ​​at the drilling location and the measured coal seam thickness was established. A seismic attribute optimization process was implemented. First, weakly correlated attributes with correlation coefficients below a set threshold were eliminated through single-well attribute-coal-thickness correlation analysis. Second, the cross-correlation matrix between attributes was calculated, and redundant attributes with cross-correlation coefficients greater than a preset value were eliminated. Finally, principal component analysis was used to perform feature dimensionality reduction on the remaining attributes. The sensitive attribute combination with the highest principal component load weight is selected to generate a spider diagram for sensitivity verification, forming the final sensitive attribute set. The selected sensitive seismic attribute data and the corresponding well point coal seam thickness samples are jointly input into the radial basis function neural network for end-to-end supervised training. This network maps the input samples to a high-dimensional space through hidden layer radial basis functions to establish a nonlinear mapping relationship between seismic attributes and coal seam thickness. The prediction error is verified using measured coal thickness data from independent verification wells. When the mean absolute error exceeds the industry allowable threshold, the attribute optimization step is returned to re-select sensitive attributes. Based on the trained neural network model, the optimized sensitive attribute set is extracted from the seismic data volume of the entire work area. After data normalization preprocessing, it is input into the network to predict the planar distribution of coal seam thickness and generate error analysis maps for geological interpretation.

2. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 1, characterized in that, The specific implementation of the principal component analysis method includes: constructing an original data matrix with coal seam thickness variation as rows and seismic attribute parameters as columns; standardizing the attribute values ​​in each column of the matrix to eliminate dimensional differences; calculating the covariance characteristics of the standardized matrix and decomposing it to obtain eigenvalues ​​and corresponding eigenvectors; selecting principal components with a cumulative contribution rate exceeding 85% by sorting the eigenvalues ​​from largest to smallest; and identifying the seismic attributes that have the most significant impact on the changes of principal components by analyzing the absolute values ​​of the eigenvector components. The screening process must simultaneously meet the dual conditions that the principal component load weight is greater than 0.7 and the correlation coefficient with coal seam thickness is greater than 0.

5.

3. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 1, characterized in that, The training process of the radial basis function neural network includes: performing range normalization on the input multidimensional sensitive seismic attribute data to eliminate the order-of-magnitude differences between different attributes; adaptively determining the hidden layer basis function centers using a nearest neighbor clustering learning algorithm, and iteratively updating the center points by dynamically calculating the Euclidean distance between new samples and existing cluster centers. When the minimum distance is less than the Gaussian function width threshold, the sample is assigned to the nearest neighbor category and the category output mean is updated; otherwise, a new cluster center is created; adaptively optimizing the basis function width parameter based on the prediction error function value, and re-clustering by iteratively reducing the width threshold until the error converges to a set range; the output layer adopts a linear weighted summation calculation method, and its weight value is the set of the output means of each category of samples, ultimately forming a continuous mapping function from the seismic attribute space to the coal seam thickness.

4. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 3, characterized in that, The dynamic update mechanism of the nearest neighbor clustering learning algorithm is as follows: the first training sample is used as the initial cluster center to establish the first hidden layer node. For each subsequent new sample, the minimum Euclidean distance from it to all existing cluster centers is calculated. If the distance is less than the current Gaussian function width threshold, the sample is assigned to the corresponding category. At the same time, the cumulative value of the output vector and the sample counter of the category are updated and the weights are recalculated. If the minimum distance exceeds the width threshold, the sample is used as the new cluster center to create a hidden layer node. The total number of hidden layer nodes in the network increases adaptively with the complexity of the sample space distribution. The output weight of each node is always the arithmetic mean of the measured coal seam thickness of all samples in the category.

5. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 3, characterized in that, The basis function width optimization adopts an error feedback control strategy: the initial Gaussian function width is set to 1 / 2 of the maximum distance in the training sample space, the sum of squared prediction errors of all training samples under the current parameters is calculated, and when the error value is greater than the convergence threshold, the width parameter is reduced according to the exponential decay rule and clustering learning is re-executed. The iterative process continues until the error function value is less than the set threshold or the maximum number of iterations is reached. The final width parameter enables the network to achieve the optimal balance between prediction accuracy and computational efficiency.

6. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 1, characterized in that, The seismic forward modeling employs a parametric wedge geological model, which systematically characterizes the continuous variation of coal seam thickness from 0.5 meters to 10 meters and covers nine typical roof and floor lithological combinations: both roof and floor are mudstone, both are limestone, both are sandstone, roof is mudstone and floor is limestone, roof is mudstone and floor is sandstone, roof is limestone and floor is mudstone, roof is limestone and floor is sandstone, roof is sandstone and floor is mudstone, roof is sandstone and floor is mudstone, and roof is sandstone and floor is limestone. By analyzing the response sensitivity of seismic attributes to coal seam thickness under different lithological combinations through forward modeling and synthesizing seismic records, a theoretical basis is provided for the optimal selection of attributes in actual work areas.

7. The method for predicting coal seam thickness based on radial basis function neural networks according to claim 1, characterized in that, The verification mechanism is implemented throughout the entire process: in the attribute optimization stage, the discriminative power of sensitive attribute combinations is analyzed through spider diagram visualization; in the neural network training stage, 20% of well points are reserved as test wells to monitor overfitting, requiring that the average relative error between the training set and the test set be less than 15%; in the work area prediction stage, cross-validation is used, and when the absolute error between the newly added drilling data and the prediction result exceeds 0.8 meters, the model optimization process is automatically triggered to ensure that the prediction results meet the accuracy requirements for coalbed methane exploration and development.

Images