Reservoir shale content prediction method based on machine learning

Abstract

The invention relates to a reservoir shale content prediction method based on machine learning. The reservoir shale content prediction method comprises the following steps: constructing a machine learning prediction model; inputting the main control factors of the shale content into the machine learning prediction model, and training the machine learning prediction model; predicting the shale content by using the trained machine learning prediction model; the main control factors are Depth, SP, CNL, GR, RT and lithology. According to the method, the shale content is predicted by adopting a machine learning method, so that the efficiency is high, the accuracy is improved, and favorable support can be provided for oil-gas exploration and development; according to the method, the precision weighted dynamic committee prediction model is established, the model integrates four network structures of BP, LightGBM, GRNN and Dilateformer, the prediction precision of the model can be further improved, the model can more flexibly adapt to the geological condition of a research area compared with a single model, and the generalization of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a method for predicting reservoir clay content based on machine learning. Background Technology

[0002] Accurate prediction of reservoir geological parameters can provide a reliable basis for reservoir evaluation. Clay content characterizes the content of silt and clay in sediments, and this parameter is usually closely related to lithological characteristics. Studies have shown a correlation between clay content and reservoir properties; generally, higher clay content leads to poorer properties, resulting in increased heterogeneity within the reservoir, affecting fluid flow paths, making it difficult to form an effective fracture network, and leading to poor fracturing performance.

[0003] For Block M in the Junggar Basin, the strata are mainly lacustrine deposits with abundant interbedded conglomerate and sandstone-mudstone. This sedimentary environment leads to significant heterogeneity in reservoir properties. The physical properties of different rock types in this block vary considerably, and the region has undergone multiple tectonic movements, resulting in a complex tectonic pattern with localized faults and folds, further complicating reservoir evaluation and productivity prediction. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a machine learning-based method for predicting reservoir clay content.

[0005] This invention is achieved through the following technical solution: The reservoir clay content prediction method based on machine learning provided by this invention includes the following steps: Construct a machine learning prediction model; input the main controlling factors of mud content into the machine learning prediction model and train the machine learning prediction model; use the trained machine learning prediction model to predict mud content.

[0006] Preferably, the main controlling factors are Depth, SP, CNL, GR, RT, and lithology.

[0007] Optionally, the following steps are also included before model training: Obtain the raw dataset, which includes clay content, GR, AC, CNL, DEN, RT, RXO, SP, CALC, CALI, depth, and lithological labels; Data augmentation was performed on clay content, GR, AC, CNL, DEN, RT, RXO, SP, CALC, CALI, depth, and lithological labels; Identify the main controlling factors affecting mud content.

[0008] Preferably, the original dataset is augmented using a CWGAN-GP network.

[0009] Optionally, the CWGAN-GP data augmentation method includes the following steps: in the discriminator and generator models of WGAN-GP, the lithological label, skewness, and kurtosis corresponding to the sample are used as constraints; the lithological label and the kurtosis and skewness of the dataset are used as condition vectors input into the CWGAN-GP model to assist in training.

[0010] Optionally, the machine learning prediction model is one of BP neural network, LightGBM, GRNN, or Dilateformer.

[0011] Preferably, the machine learning prediction model is a dynamic committee ensemble model.

[0012] Preferably, four expert models—BP neural network, LightGBM, GRNN, and Dilateformer—are used to construct the committee machine prediction model.

[0013] Optionally, the method for constructing a dynamic committee integration model includes the following steps: Fuzzy C-means clustering is used as a gate network for dynamic integration to divide the data base into subsets, dividing the dataset into K clusters, and the optimal matching relationship between each expert model and the subset is determined by the fuzzy C-means clustering method. Assign weights to the expert models within each cluster; For each sample in the dataset, calculate its membership degree in each cluster; Based on the membership degree of the sample in each cluster and the model weight of the corresponding cluster, the membership degree of the sample and the model weight value of the corresponding cluster are multiplied together to obtain the weight allocation result of each model to each sample in each cluster. Furthermore, the weight values ​​of each model to a certain sample in all clusters are accumulated to obtain the dynamic weight of the expert model for that sample. The dynamic weights of each expert model on the samples are assigned, and the prediction results of each expert model are weighted and integrated to form a dynamic committee ensemble model.

[0014] Compared with the prior art, this application has at least the following beneficial effects: 1. This application uses machine learning methods to predict mud content, which is not only efficient but also improves accuracy, providing favorable support for oil and gas exploration and development. 2. This application screens key control factors, which can effectively eliminate irrelevant or redundant input features, preventing the curse of dimensionality caused by too many features and limited sample size, thus avoiding overfitting during model training. Furthermore, by screening key control factors, the complexity of hyperparameter optimization during model training can be effectively simplified, making it easier to find the optimal solution. At the same time, it clearly quantifies the influence of each logging parameter on the geological parameters of the target reservoir, helping geologists understand the dominant factors in evaluating each reservoir geological parameter.

[0015] 3. In one optional scheme of this application, a precision-weighted dynamic committee prediction model is established. This model integrates four network structures: BP, LightGBM, GRNN, and Dilateformer, which can further improve the prediction accuracy of the model. Compared with a single model, it can more flexibly adapt to the geological conditions of the study area and improve the generalization of the model. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is an example diagram of the LightGBM histogram algorithm in the embodiment; Figure 2 The diagram shows the BP neural network structure in the embodiment. Figure 3 The GRNN network structure is shown in the example. Figure 4 This is a schematic diagram of multi-scale extended attention MSDA in the embodiment; Figure 5 This is a structural diagram of the multi-scale dilated attention mechanism in the embodiment; Figure 6 This is a flowchart of the CWGAN-GP training process in the embodiment; Figure 7 This is a comparison chart of kurtosis before and after the increase in mud content in the example; Figure 8 This is a comparison chart of statistical indicators before and after the enhancement of mud content in the examples; Figure 9 This is a clustering result diagram of lithological tags for mud content in the examples; Figure 10 This is a graph showing the correlation between mud content in the examples; Figure 11 This is a flowchart of the training process for the mud content prediction model in the embodiment; Figure 12 This is a comparison chart of the predicted mud content of each model in the embodiments; Figure 13 This is a flowchart of the accuracy-weighted dynamic committee ensemble model prediction process in the embodiment; Figure 14 This is a framework diagram for predicting clay content based on a precision-weighted dynamic committee ensemble model in the implementation example; Figure 15 This is a graph showing the predicted reservoir clay content results from the dynamic committee integrated model in the embodiment. Figure 16 This is a graph showing the predicted mud content before data enhancement in the example. Figure 17 This is a comprehensive interpretation diagram of reservoir geological parameters in well X1 in the embodiment; Figure 18 This is a comprehensive interpretation diagram of reservoir geological parameters in well X2 in the example. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. It should also be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0020] Example 1: The machine learning-based method for predicting reservoir clay content disclosed in this example includes the following steps: 1. In this embodiment, the reservoir geological parameter calculation model established based on core data and well logging response characteristic analysis in the same study area is used to calculate and compare the clay content collected in this study.

[0021] Since neutrons, density, and clay content are all correlated, a multiple regression method is used to calculate clay content based on neutrons and density. : ; In the above formula, The content of clay, To compensate for neutrons, This represents the density of the formation.

[0022] 2. Construct machine learning prediction models (1) LIGHTGBM is a machine learning algorithm based on Gradient Boosting Tree (GBDT). It fits new decision trees sequentially along the negative direction of the decision tree loss function. Using a histogram algorithm, each continuous feature value is divided into a finite number of histograms, such as... Figure 1 As shown, the optimal split point is selected from the candidate set of each feature, and the corresponding bin is selected as the gain of the split point to accelerate model training. Furthermore, in Gradient One-Sided Sampling (GOSS), information gain is calculated only for samples with large gradients, and only a portion of samples with smaller gradients are randomly retained, thus preserving the distribution of the original data as much as possible. By bundling and merging mutually exclusive features, training efficiency is improved, ensuring good generalization ability while enhancing model accuracy.

[0023] Compared to traditional machine learning models such as linear regression and SVM, LIGHTGBM is more flexible, better able to capture complex mapping relationships in data, and can automatically select features from parameters and logging data that affect clay content prediction, reducing the need for manual intervention. Compared to convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM), LIGHTGBM is more lightweight and suitable for resource-constrained environments. Deep learning models often have a black-box nature, with internal decision-making processes difficult to interpret, while LIGHTGBM can better reveal the influence of internal features on reservoir geological parameters and can more intuitively analyze feature importance to guide further logging evaluation studies.

[0024] (2) The basic structure of a BP neural network consists of an input layer, hidden layers, and an output layer. For example... Figure 2 As shown, BP neural networks possess powerful nonlinear modeling capabilities, enabling them to capture complex nonlinear relationships in reservoir geological parameters. This is extremely useful for understanding and predicting complex geological features.

[0025] (3) GRNN neural network, also known as generalized regression neural network. This network is trained based on nonparametric kernel regression. By calculating the correlation density function between the independent and dependent variables from the training samples, the regression value of the dependent variable relative to the independent variable is obtained. The GRNN neural network consists of four layers of neurons. The first layer is the input layer, and the number of neurons in this layer is the same as the dimension of the input vector in the training samples. That is, each neuron can be regarded as a simple distributed neural network. The second layer is the pattern layer, and the number of neurons in this layer is the same as the number of training samples. The summation layer sums the results of all training samples, and the output layer outputs the prediction results. The structure diagram of the GRNN neural network is shown below. Figure 3 As shown.

[0026] Because GRNN can quickly adapt to data distribution without requiring complex parameter tuning, it is suitable for complex and variable well logging and geological data. Furthermore, by adjusting the smoothing factor of the network, it can provide smooth prediction results, making it suitable for prediction tasks involving continuous variables with deep sequential characteristics. Compared to traditional machine learning methods, GRNN is more flexible, adept at capturing complex relationships in the data, and its parameter tuning is simpler and computationally more efficient than deep learning models.

[0027] Therefore, this embodiment selects the GRNN neural network as a prediction model to predict the mud content.

[0028] (4) Dilateformer is a multi-scale dilated attention mechanism neural network. In general convolutional neural network structures, sliding windows can only handle features at a fixed scale. However, the Dilateformer network introduces dilated convolution operations in the multi-head attention mechanism module to capture the dependencies between features at different scales, thereby better understanding the complex structure and patterns of input features. Furthermore, through sparse selection, it improves the model's operating efficiency. The Dilateformer model generally presents a four-layer pyramid structure, as shown below. Figure 5 As shown, the first two layers are Multi-Scale Extended Attention (MSDA). These two layers can be understood as performing a sliding window operation within a multi-head attention mechanism. The execution process is as follows: Figure 4 As shown. The last two layers are standard Multi-Head Self-Attention (MHSA). Overlapping markers are used in each layer, with zero-padding achieved through multiple overlapping 3×3 convolutional modules. The output resolution is adjusted by alternately controlling the stride of the convolutional kernel to 1 or 2. A convolutional overlap downsampling unit with an overlap kernel size of 3 and a stride of 2 is used to adapt each layer to inputs at different resolutions. Furthermore, whenever feature information is input to MSDA or MHSA, conditional embedding (CPE) is used to positionally encode the input and depth logging curves, enabling the model to better adapt to feature variations at different depth locations.

[0029] In MSDA, the Sliding Window Extended Attention (SWDA) operation is used. This involves sparsely selecting the keys and values ​​of each header within a sliding window centered on the red query patch, and then performing self-attention calculation on these representative patches. The calculation formula is as follows: ; In the above formula, X is the calculated self-attention score; Q, K, and V represent the query, key, and value matrices, respectively, with each row of each matrix representing a query / key / value feature vector; r is the inflation rate corresponding to different sliding windows; and in the original feature map... When querying at a location, SWDA will use... Self-attention computation is performed by sparsely selecting keys and values ​​within a sliding window of size w×w centered on the network. This represents the x-coordinate value corresponding to the center point of the sliding window. This represents the ordinate value corresponding to the center point of the sliding window. The dimensions are the length and width of the rectangular sliding window.

[0030] Because reservoir geological parameters typically exhibit complex depth sequence relationships, the multi-head attention mechanism in the Dilateformer model effectively captures long-distance dependencies between well logging data, which helps in understanding the potential relationships between different locations within the reservoir. Furthermore, by adjusting the dilation rate, dilated convolution can extract features at different scales, expanding the model's receptive field and enabling it to cover a larger depth sequence region—something traditional machine learning models (such as linear regression and support vector machines) struggle to achieve. Unlike LSTM and RNN models, Dilateformer does not rely on the sequential information of the depth sequence, thus allowing for more flexible processing of well logging data with longer depth sequences.

[0031] 3. Flowchart of Machine Learning-Based Reservoir Slurry Content Prediction Method 3.1 Introduction to the dataset and data preprocessing This embodiment uses nine common logging parameters I, along with depth and lithology labels, to calculate clay content.

[0032] Nine common logging parameters are GR, AC, CNL, DEN, RT, RXO, SP, CALC, and CALI. Among them, GR represents natural gamma, reflecting the total content of radioactive elements in the formation; AC represents acoustic transit time, which is a measure of the difference between the speed of sound propagation in rock and the speed of sound propagation in a standard medium, often used to identify the lithology and thickness of formations and to identify oil and gas-bearing layers; CNL represents compensated neutrons, which refers to the thermal neutron flux caused by a neutron source measured along the well profile, used to estimate the hydrogen index and porosity of the formation; DEN represents formation density, which refers to the density value of different formations corresponding to different lithologies, used to determine the thickness and lithology of the formation; RT represents deep resistivity, which refers to the resistivity of the formation before drilling fluid intrusion, used to determine the oil-bearing nature of the formation; RXO represents flush zone resistivity, which refers to the resistivity in the flush zone after drilling fluid intrusion, used to determine the degree of influence of drilling fluid on the formation; SP represents spontaneous potential, which refers to the potential difference generated by the resistivity of drilling fluid adjacent to porous formations, used to delineate permeable layers, determine lithology and formation water properties; CALC represents well diameter difference, which refers to the difference between the actual well diameter and the standard well diameter. Used to identify wellbore irregularities and provide a basis for formation evaluation; CALI stands for wellbore diameter, which usually refers to the diameter of the wellbore and is used to determine lithology, check casing condition, etc.

[0033] The experimental dataset for the machine learning-based reservoir clay content prediction method in this embodiment comes from measured core samples of the Wuerhe Formation in Block M of the study area. Clay content data were recorded in 137 samples. Combined with lithological labels and well logging parameters at corresponding depth points, this constitutes the reservoir geological parameter sample dataset for this embodiment. Well C was used for clay content analysis. The predicted data is shown in Table 1.

[0034] Table 1: Prediction Parameter Table

[0035] The dataset used to predict mud content was divided into training and testing sets in a 7:3 ratio. Due to measurement errors and differences in evaluation metrics during data collection, a maximum-minimum normalization method was used to ensure that the distribution of each feature data was within a reasonable range.

[0036] 3.2 Improved CWGAN-GP Data Augmentation Process like Figure 6 As shown, the CWGAN-GP generative adversarial network (GAN) consists of two neural networks: a generator and a discriminator. Through multiple rounds of adversarial training, it generates distributions similar to the original samples. WGAN-GP is an improved GAN model, using JS divergence to measure the difference between distributions. The discriminator must satisfy the Lipsehitz condition to ensure the effectiveness of the Wasserstein distance and the stability of training. The goal of network training is to maximize the discriminator's probability of judging real samples (D(x)) to better distinguish between real and generated samples. Minimizing the generator loss function makes the generated samples closer to real samples, thus deceiving the discriminator. To further improve the quality of sample generation, this embodiment uses conditional constraints and feature extraction capabilities. However, if the constraints only contain the sample label information, it is difficult to learn the inherent distribution patterns of the feature data during adversarial training, limiting further optimization iterations of the generator. Therefore, in the discriminator and generator models based on the original WGAN-GP model, in addition to using the lithology label corresponding to the sample as a constraint, additional dimensionless skewness and kurtosis characteristic indices are added to achieve conditional generation, which constitutes the CWGAN-GP network structure in this embodiment.

[0037] The formulas for calculating skewness and kurtosis are shown in Equations (1) and (2) to stably reflect and limit the distribution of geological and engineering fracturing data.

[0038] (1); (2); In the above formula, K represents the kurtosis of the dataset, and S represents the skewness of the dataset. This represents a sample in the dataset, where Std is the standard deviation, M is the sample mean, and N is the total number of samples in the dataset.

[0039] Two conditional values, derived from the formulas for kurtosis and skewness, are concatenated with the sample lithology label L to obtain a conditional vector C=[L,K,S]. This conditional vector is then input into the WGAN-GP model to aid training. The generator generates new samples, and the discriminator distinguishes between the generated and real samples. The parameters of the generator and discriminator are updated by calculating the loss function.

[0040] The loss function of the CGAN-GP discriminator is: (3); In the above formula, Z represents the discriminator loss function; Z represents the real sample. The data distribution represents the real samples; x represents the samples generated by the generator; y represents the conditional constraints added during the training of the discriminator. The prior distribution representing the noise; Linear interpolation representing real samples and generated samples; The data distribution representing the linear interpolation samples; The expected score of the discriminator on the real data. The expected score of the discriminator on the generated data; This is the gradient penalty term for the discriminator. λ The coefficient of the gradient penalty term, To introduce a gradient penalty term for linear interpolation, the gradient penalty term is added to the generated linear interpolation samples to ensure that the discriminator applies the gradient penalty uniformly throughout the data space. This ensures that the discriminator satisfies the Lipschitz condition across the entire data distribution, which helps the discriminator to have a smooth gradient not only on real and generated data points, but on all points, thus effectively preventing gradient vanishing or gradient exploding.

[0041] The loss function of the generator is: (4); In the above formula, represents the discriminator loss function; x represents the samples generated by the generator; y represents the constraints added during the generator training process. Let $\frac{ ...

[0042] Data Augmentation Results: During the data augmentation process of the original dataset of mud content, the average sample sampling rate (ASOR), i.e., the average proportion of the generated data of each feature falling within the value range of the original data, was used to evaluate the similarity between the distribution of the original sample and the generated sample. The ASOR calculation formula is shown in (5)-(7): (5); (6); (7); In the above formula, ASOR is the average sample sampling rate. To generate samples for the adversarial network, n is the number of samples generated, [min, max] is the value range of the original data, d is the total number of features in each sample dataset, i is the index of each sample, and j is the index of each feature. The sampling rate corresponds to different features.

[0043] The hyperparameter settings for CWGAN-GP during data augmentation are shown in Table 2. Typically, the iteration rate for both the generator and discriminator is set to 5 by default to improve the generator's training efficiency and reduce the risk of pattern collapse during training.

[0044] Table 2: CWGAN-GP Model Parameter Settings Table

[0045] CWGAN-GP was used to augment the clay content dataset, and the results of GAN network augmentation on three reservoir geological parameter datasets were compared. An ASOR threshold of 95% was set to determine the number of samples in the generated datasets for each reservoir geological parameter; ultimately, the number of generated samples for clay content was determined to be 3268. Figure 7 The charts show a comparison of the changes in skewness and kurtosis of various features of the mud content data after data enhancement using CWGAN-GP and GAN networks, respectively, with the original data. It can be seen that, compared to the data enhanced by GAN, CWGAN-GP accurately preserves the positive or negative kurtosis of the original data and can more accurately fit the kurtosis values ​​of each feature of the original data, maximizing the diversity of the original data while ensuring that each type of data feature is learned.

[0046] To further verify the learning performance of the adversarial network in this embodiment on the distribution of the original reservoir geological parameter dataset, Figure 8 The statistical indicators of various characteristics before and after the enhancement of mud content data were compared.

[0047] Therefore, although GANs are closer to the original data than CWGAN-GP in some features, GANs' ability to learn the distribution of the original data is not very stable, and the quality of the generated data is not as good as that of CWGAN-GP. The data generated by CWGAN-GP is relatively close to the corresponding features of the original data in most aspects. However, there are still some significant differences in some features. For example, in the mud content generated sample set, the mean and median of the RT and RXO generated samples are significantly larger. This is mainly because the values ​​and fluctuations of deep and shallow resistivity in the original sample set are large. In the mud content dataset, the GR and AC of the generated samples have significant errors in both the mean and median. This is because in the reservoir geological parameter dataset, most index data are small, while the mean values ​​of GR and AC are both above 68, belonging to the large value category. Large value features account for a small proportion of the total features, resulting in CWGAN-GP being able to learn the overall distribution trend of the data, but its learning ability for large value features is poor.

[0048] Maximum homogeneity difference (MMD) is a popular criterion for measuring the difference in distributions of similar data in mathematical statistics tasks. Two different data distributions are mapped to a regenerated Hilbert space H through a mapping function. In this high-dimensional space, the distance between the probability distributions of the original sample and the generated sample is evaluated by the maximum homogeneity difference, calculated as (8): (8); In the above formula, and These represent the reservoir geological parameter datasets before and after data augmentation. and They represent and The first in One sample, and Indicates the number of data points in the two domain datasets. It is a mapping function that maps data from a low-dimensional space to a high-dimensional space.

[0049] Table 3 shows the MMD values ​​of clay content distribution before and after CWGAN-GP and GAN data augmentation. Compared with GAN networks, the probability distribution of CWGAN-GP generated samples is closer to that of real samples. The MMD values ​​of the three parameters are all between 0.25 and 0.5, indicating that the samples generated by CWGAN-GP can be used as extended samples of the original core samples to increase the number of training samples, which is beneficial to improving the training ability of the subsequent model and improving the generalization performance of the model.

[0050] Table 3: Comparison of MMD values ​​before and after mud content enhancement

[0051] The t-SNE visualization tool was used to verify the clustering effect of CWGAN-GP on lithology labels before and after augmentation of three reservoir geological parameter datasets. The clustering results are as follows: Figure 9 As shown, after CWGAN-GP data augmentation, the mud content dataset can further improve the clustering trend of lithological labels of the same category, effectively preserve the separation characteristics between different lithologies, and also learn that there are similar overlapping relationships between different lithological samples as in the original samples.

[0052] It is evident that although GANs are closer to the original data than CWGAN-GP in some features, the distribution of the GAN-synthesized data differs significantly from the original data for most features. Overall, GANs' ability to learn the distribution of the original data is less stable, and the quality of the generated data is inferior to that of CWGAN-GP. The data generated by CWGAN-GP, on the other hand, is quite close to the corresponding features of the original data in most aspects.

[0053] 4. Analysis of Controlling Factors The Pearson correlation coefficient is a linear correlation function used to analyze the linear correlation between two features. Suppose there exist two feature sets X and Y: , The formula for calculating the Pearson correlation coefficient is: (9); In the above formula, and represents the average values ​​of x and y, respectively, and r represents the Pearson correlation coefficient.

[0054] Kendall's rank correlation coefficient is used to measure the consistency of changes in the rank order of two variables. For any two adjacent data points (Xi, Yi) and (Xj, Yj) at positions i and j in two feature sets X and Y, if the relationship between Xi and Xj is consistent with the relationship between Yi and Yj, then this set of data points is called a consistent pair; otherwise, it is called an inconsistent pair. Based on the ratio of consistent pairs to inconsistent pairs in the two sets of data, Kendall's rank correlation coefficient is calculated to measure the consistency of changes in the rank order of the two sets of variables. The calculation formula is shown in (10).

[0055] (10); In the above formula, Kendall's correlation coefficient; The number of consistent pairs; The number of inconsistent pairs; The total logarithm; Let X be the number of parallel pairs. Let Y be the number of parallel pairs.

[0056] The Spearman correlation coefficient is used to measure the monotonic correlation between two variables. Unlike the Pearson correlation coefficient, which requires the data to follow a normal distribution, the Spearman correlation coefficient does not have strict requirements on the data distribution. When calculating the Spearman correlation coefficient of two sets of data, the data in the two feature sets X and Y are first assigned ranks in descending order. For each data point i, the difference between its ranks in X and Y is calculated, as shown in formula (11). The Spearman correlation coefficient of the two sets of data is then calculated based on the rank difference, as shown in formula (12).

[0057] (11); (12); In the above formula, , For data points Rank the data in the two sets; For data The difference in rank between the two sets of data.

[0058] The range of all three correlation coefficients is [-1, 1], where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no correlation. In practical research, the strength of the correlation is usually graded based on the absolute value of the correlation coefficient. Specifically, 0-0.3 represents a weak or negligible correlation; 0.3-0.5 represents a low correlation; 0.5-0.9 represents a moderate to high correlation between the two variables; and values ​​above 0.9 represent an extremely strong correlation between the variables.

[0059] The linear and nonlinear correlations between influencing factors and clay content were calculated using Pearson correlation coefficient and two rank correlation coefficients from Kendall's and Spearman, respectively. A weighted average of the results from the three correlation coefficient calculation methods was then used to combine their performance under different data characteristics, resulting in a more comprehensive and robust correlation assessment. Absolute values ​​were used for all correlation parameters during the calculation. A larger mean absolute value of the weighted correlation coefficient of a parameter with clay content indicates a greater influence of that parameter on clay content prediction.

[0060] The standard for screening the main controlling factors of each reservoir geological parameter is: if, for the mud content, the weighted average of the absolute value of the correlation coefficient of one of the influencing parameters is higher than the average of the weighted average of the absolute values ​​of the correlation coefficients of all influencing parameters under the mud content, then the influencing parameter is considered to be the main controlling factor for predicting the mud content.

[0061] Figure 10 The graph shows the correlation between clay content and corresponding influencing factors. Calculation results indicate that Depth, SP, CNL, GR, RT, and lithology have a significant impact on clay content prediction. These parameters are selected as the main controlling factors in the corresponding machine learning-based reservoir clay content prediction methods.

[0062] 5. Evaluation Indicators for the Computational Model This embodiment uses the coefficient of determination (COP). The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to measure the accuracy of each reservoir geological prediction model. The MAE represents the average of the sum of the absolute values ​​of all prediction errors, and the MSE represents the square root of the mean of the squared prediction errors for all data. The formulas for calculating these evaluation metrics are as follows: (13) (14) (15) in, This represents the actual values ​​of reservoir geological parameters. ` represents the average value of the actual values ​​of reservoir geological parameters. Represents the predicted values ​​of reservoir geological parameters. The sample number. This represents the total number of samples in the dataset.

[0063] 6. Algorithm Simulation Process 6.1 Model Training Process The process of predicting reservoir clay content based on four machine learning algorithms can be divided into the following steps: (1) Data anomalies and missing values ​​were handled for the collected mud content dataset. The CWGAN-GP model with added constraints was used to augment the mud content dataset. The rationality of the features and category labels of the augmented samples was verified by statistical index analysis and the t-SNE visualization tool. The dataset was normalized to eliminate the dimensional differences between different features.

[0064] (2) After data augmentation, the mud content data was screened for controlling factors and then divided into training and test sets. The training set data was input into four regression models: BP, Grnn, LightGBM, and Dilateformer. The Bayesian optimization algorithm was used to iteratively search the hyperparameter search space to determine the optimal hyperparameter combination. After determining the optimal hyperparameter combination, the training effect was evaluated using indicators such as the coefficient of determination (R2) and root mean square error (RMSE). The generalization ability of each model was evaluated on the test set data. To verify the effectiveness of data augmentation and the generalization ability of the model, the original sample data corresponding to each parameter was applied to the trained optimal model and the model performance was evaluated. The model training process is as follows: Figure 11 As shown.

[0065] 6.2 Introduction to Model Hyperparameters (1) The main parameters and optimization range settings in the BP neural network prediction model are as follows: Hidden_layers_sizes: This refers to the number of hidden layer neurons in the network. Setting this parameter affects the model's expressive power. Setting too many layers increases training time and can lead to overfitting, while setting too few layers can prevent the model from fully capturing the feature representations between data points, potentially resulting in underfitting. In this embodiment, the hyperparameter optimization range for the BP neural network is set to (5, 30).

[0066] Learning_rate: In this embodiment, the learning rate hyperparameter optimization range of the BP neural network is set to (0.001, 0.1).

[0067] Maximum number of iterations: In this embodiment, the hyperparameter range for the maximum number of iterations is set to (10, 500).

[0068] Regularization coefficient: This parameter is used to adjust the complexity of the model during training. Setting it too high will lead to oversimplification of the model, resulting in underfitting; setting it too low will result in insufficient model complexity, leading to overfitting. In this embodiment, the optimization range of the regularization coefficient hyperparameter is set to (0.0001, 0.01).

[0069] (2) In the LightGBM model, the main hyperparameters and optimization range are set as follows: N_estimators: This refers to the number of decision trees, which affects the model's complexity. Too many trees may lead to overfitting, while too few trees may lead to underfitting, preventing the model from fully capturing data features. In this embodiment, the decision tree hyperparameter optimization range is set to (0, 300).

[0070] Max Depth: This limits the maximum growth depth of each decision tree. Its purpose is to regulate the complexity of the model training process to prevent overfitting. In this embodiment, the maximum growth depth hyperparameter is set to (1, 10).

[0071] Min Child Samples: This refers to the minimum number of leaf nodes in the tree. This parameter limits the growth of the decision tree and the complexity of the model. In this embodiment, the optimization range for the minimum number of leaf nodes is set to (1, 20).

[0072] Lambda_L1 and Lambda_L2 refer to the large gradient sample sampling rate and the large gradient random sample sampling rate, respectively. These two parameters allow the model to randomly select a subset of samples and features for training during a single iteration. Random sampling improves the model's generalization ability. The optimization range for both Lambda_L1 and Lambda_L2 is set to (0, 1).

[0073] The prediction results of the GRNN neural network are related to the smoothing factor. Setting this parameter too high will cause the prediction results to approximate the mean of the sample data. Setting it too low will cause the model to be overly sensitive to local changes in the input samples, resulting in an unsmooth prediction curve. In this embodiment, the smoothing factor is optimized within a range of (0.1, 10).

[0074] (3) In the Dilateformer model, the main hyperparameters and optimization range are set as follows: Droupout refers to a technique used in neural networks to randomly disable neurons during training. Specifically, during forward propagation, a certain proportion of neurons are deactivated, effectively preventing overfitting and improving the model's generalization ability. In this embodiment, the optimization range for Droupout is set to (0.1, 1). The optimization ranges for other hyperparameters are as follows: Hidden_layers_sizes: (5, 30), Num_Layers: (0, 5), Learning_rate: (0.001, 0.1).

[0075] In the CWGAN-GP network, besides appropriately setting the generator and discriminator iteration counts, learning rates, and iteration counts, two additional hyperparameters need to be considered: the gradient decay factor and the squared gradient decay factor. These two parameters are used to calculate the gradient norm of the linear interpolation point between real and generated samples, thus controlling the penalty applied to the gradient norm during model training. Typically, these two parameters are set within the range of (0.1, 10) to ensure the gradient norm is as close to 1 as possible, effectively preventing gradient explosion.

[0076] 7. Prediction of mud content The enhanced clay content dataset was divided into training and testing sets in a 7:3 ratio. Four models were used to predict clay content. The prediction accuracy was 0.86 for BP neural network, 0.88 for LightGBM, 0.89 for GRNN, and 0.89 for Dilateformer. The prediction results are as follows: Figure 12 (a)-(d) are shown. Table 4 lists the optimal hyperparameters for each model when predicting mud content.

[0077] Table 4: Results of Hyperparameter Optimization

[0078] Table 5 shows the best performance indicators for predicting clay content on the test set for four machine learning models after hyperparameter optimization. It can be seen that the Dilateformer neural network exhibits more accurate prediction performance when predicting clay content, with a slightly better prediction accuracy than the other models, while the GRNN neural network also shows high prediction accuracy. That is, these two neural network models can accurately fit the nonlinear correlation between clay content and its influencing factors based on complex internal nonlinear functions.

[0079] Table 5: Comparison of Evaluation Indicators for Prediction of Clay Content

[0080] Table 6 records the prediction evaluation indices (RMSE, MAE) of the four models for the clay content data of the core points. The prediction results of the four models for the clay content of the core points are compared with the RMSE and MAE calculation results obtained by the multiple regression calculation method for clay content established by previous researchers in the same study area.

[0081] Table 6: Comparison of Prediction and Evaluation Indicators for Core Data on Muddy Content

[0082] By comparing and analyzing the above evaluation indicators, it can be concluded that the prediction results of the four machine learning models are better than the calculation results of traditional methods for mud content.

[0083] 8. A machine learning-based method for predicting reservoir clay content using the improved committee model. 8.1 Committee Machine Prediction Model The committee-based machine prediction model is an ensemble neural network algorithm whose core idea is "brainstorming." It's a machine learning technique that improves overall prediction performance by combining multiple experts (i.e., individual models). Committee models are divided into static committee models and dynamic committee ensemble models. Static committee models use a fixed weighted average of the prediction results from multiple experts, assigning fixed weights to each expert model. For data with multiple different patterns or distributions, fixed weights may not fully capture the features of all patterns, making it difficult to improve the training performance of individual experts. Dynamic committee models, on the other hand, use a gate network to divide the input data into multiple sub-tasks and assign them to the corresponding expert models; then, they integrate the results from multiple different models to output the final prediction result. Fuzzy C-means clustering is typically used as the gate network for weight allocation in dynamic ensemble modeling. This method divides the data into sub-data based on the principle that the inter-class differences are sufficiently large and the intra-class differences are sufficiently small. The formula is the fuzzy C-means clustering calculation formula: (16); In the above formula, The loss function for fuzzy clustering is used to measure the quality of the current clustering results; U is the membership matrix, V is the cluster centroid, X is the input well logging data, M is the number of data entries, C is the number of clusters, and q is the fuzzy coefficient. For the i-th data point in the sample dataset, For the k-th cluster center, Let i represent the degree to which data point i belongs to cluster k, where i is the index of the base sample point and k is the index of the cluster center.

[0084] Directly combining samples and models to obtain prediction results can weaken the relationship between the model and samples to some extent. Therefore, this embodiment uses a dynamic weighted ensemble method to achieve weight allocation. First, weights are allocated to the model. For a single sample... Calculate its membership degree to each cluster. Furthermore, for the membership degree of the same sample in different clusters, the weighted sum of all membership degrees must be 1. Then, calculate the membership degree of each expert learner. For the sample final weight Based on the calculated dynamic weights, the optimal matching relationship between each expert and the subset of data is determined. The various expert models are then combined to form the final prediction model. The model training process is as follows: Figure 13 As shown.

[0085] Considering the excellent self-learning and adaptive capabilities of BP neural networks, the strength of LightGBM in handling complex relationships between features, the suitability of GRNN for approximating complex functions, and the proficiency of Dilateformer in capturing long-range dependencies between well logging data, and given the complex geological conditions and strong reservoir heterogeneity in Block M, a dynamic committee ensemble model is constructed using these four models. Specifically, the construction method of the dynamic committee ensemble model includes the following steps: This method uses fuzzy C-means clustering as a gate network for dynamic ensemble. It partitions the data base into subsets based on the principle of sufficiently large in-class and sufficiently small in-class differences. The dataset is divided into K clusters using fuzzy C-means clustering. The optimal matching relationship between each expert model and its subsets is then determined using the fuzzy C-means clustering method.

[0086] Assigning weights to each expert model: Within each cluster after partitioning, the performance of each expert learner is evaluated using the weighted mean squared error (MSE). Weights are assigned to models within each cluster based on their weighted mean squared error values; generally, models with better performance are given higher weights.

[0087] Membership degree calculation based on samples: For each sample in the dataset, calculate its membership degree in each cluster, that is, the degree to which the sample belongs to each cluster. For the same sample, the weighted sum of its membership degree values ​​in each cluster should generally be equal to 1.

[0088] A dynamic weighted ensemble method is used to assign weights to each expert model: based on the membership degree of the sample in each cluster and the model weight of the corresponding cluster, the membership degree of the sample and the model weight of the corresponding cluster are multiplied together to obtain the weight assignment result of each model to each sample in each cluster. Furthermore, the weight values ​​of each model to a certain sample in all clusters are accumulated to obtain the dynamic weight of the expert model for that sample. The dynamic weights of each expert model on the samples are assigned, and the prediction results of each expert model are weighted and integrated to form a dynamic committee ensemble model. This model can dynamically adjust the weights of each model according to different geological conditions and data characteristics to better adapt to the complex geological conditions of Block M.

[0089] The specific steps for implementing dynamic weight allocation using fuzzy C-means clustering are as follows: Step 1, Feature Clustering: Use the features of the training set to perform FCM clustering to divide the data into K clusters.

[0090] Step 2, evaluate model performance: Within each cluster, evaluate the performance of each base learner by weighted mean square error (WeightedMSE), as shown in Equation (17).

[0091] (17); In the above formula, It is a cluster The number of samples; It is an expert model For the sample The predicted value, It is a sample The true value, Let be the membership degree of sample j to cluster K.

[0092] Step 3, Weight Allocation: Based on the model's performance within each cluster Assign weights to the model for each cluster. For example, the calculation formula (18) suggests that the better the model performs, the higher its weight.

[0093] (18) In the above formula, ε is a small constant, introduced to avoid the situation where a certain model... When ε is 0, the denominator becomes 0, resulting in a division-by-zero error. Typically, ε takes values ​​much smaller than the order of magnitude of the MSE, such as when the predicted value's MSE is on the order of magnitude of... - In between, MSE is in - Random numbers are generated between the given values.

[0094] Step 4, Prediction Phase: First, for each test sample... Calculate its membership degree to each cluster, that is, the degree to which the sample belongs to each cluster. If the data is divided into K clusters, then k takes values ​​from 1 to K, and satisfies the following condition: Calculate the final weight of each model based on its membership degree and the model weight of the corresponding cluster. As shown in formula (19); the prediction results of the base learners are weighted based on dynamic weights, and the final ensemble prediction value is obtained by formula (20). .

[0095] (19); (20); By using a dynamic ensemble approach, the weights of each base expert are dynamically adjusted based on the characteristics of the input samples, the optimal matching relationship between each expert and the subset is determined, and more accurate prediction results are generated.

[0096] When predicting porosity and permeability, the ensemble models that contribute the most to the prediction results are Dilateformer, GRNN, Dilateformer, and BP, respectively. Table 8 shows the dynamic weight allocation for each model. The overall framework of the improved accuracy-weighted dynamic committee ensemble model for predicting porosity and permeability is as follows: Figure 14 As shown.

[0097] Table 7: Model Weights for Predicting Clay Content 8.2 Model Prediction Results like Figure 15 As shown, the results of the dynamic committee ensemble model predicting clay content are presented in Table 8. Table 8 records the performance evaluation indicators of the ensemble model in predicting clay content. Table 8: Evaluation Indicators for Predicted Clay Content Using Dynamic Committee Integrated Model Figure 16 The prediction results of the ensemble learning model for the original core data points with mud content were visualized.

[0098] As can be seen, the dynamic committee ensemble model constructed in this embodiment significantly improves the prediction performance of mud content compared to the single model. The model maintains a good fit between the predicted and actual mud content values, with a relatively uniform error distribution.

[0099] 8.3 Single-well processing interpretation application examples Based on the dynamic committee integration model established above, comprehensive well logging interpretation of reservoir geological parameters was carried out on X1 and X2, two verification wells in block M that participated in model training. Figure 17 , Figure 18 The results show that the prediction curves of the dynamic committee integrated model for clay content are in high agreement with the core data curves. This indicates that the model still has high predictive performance even with noise and uncertainty in the data.

[0100] The dynamic committee ensemble model established in this embodiment integrates four network structures: BP, LightGBM, GRNN, and Dilateformer. Based on this model, the clay content is predicted, and the results show a high correlation coefficient between the predicted clay content and the actual prediction. The accuracy reached 0.95, indicating that the dynamic committee ensemble model can further improve the prediction accuracy of the model, and can adapt to the geological conditions of the study area more flexibly than the single model, thus improving the generalization ability of the model.

[0101] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting reservoir shale content based on machine learning, characterized in that, The method comprises the following steps: constructing a machine learning prediction model; inputting the main control factors of shale content into the machine learning prediction model, and training the machine learning prediction model; using the trained machine learning prediction model to predict the shale content.

2. The machine learning based reservoir shale content prediction method of claim 1, wherein, The main control factors are depth, SP, CNL, GR, RT and lithology.

3. The machine learning based reservoir shale content prediction method of claim 1, wherein, The method for determining the main control factors comprises the following steps: calculating the linear and nonlinear correlation degrees between the shale content and the corresponding influence parameters by using Pearson correlation coefficients and Kendall's and Spearman's two rank correlation coefficients; taking the absolute values of the three correlation coefficient calculation values and performing weighted average; if the absolute value weighted average value of the correlation coefficient of one of the influence parameters is higher than the average value of the absolute value weighted average values of the correlation coefficients of all the influence parameters, the influence parameter is considered as the main control factor.

4. The machine learning based reservoir shale content prediction method of any one of claims 1-3, wherein, Before the model training, the following steps are further included: obtaining an original data set, wherein the original data comprises shale content, GR, AC, CNL, DEN, RT, RXO, SP, CALC, CALI, depth and lithology label; performing data enhancement on the shale content, GR, AC, CNL, DEN, RT, RXO, SP, CALC, CALI, depth and lithology label; determining the main control factors affecting the shale content.

5. The machine learning based reservoir shale content prediction method of claim 4, wherein, The CWGAN-GP network is used to perform data enhancement on the original data set.

6. The machine learning based reservoir shale content prediction method of claim 5, wherein, The method for data enhancement by using the CWGAN-GP comprises the following steps: in the discriminator and generator models of the WGAN-GP, the lithology label, skewness and kurtosis corresponding to the sample are taken as constraint conditions; the lithology label and the skewness and kurtosis of the data set are taken as a condition vector and input into the CWGAN-GP model for auxiliary training.

7. The machine learning based reservoir shale content prediction method of claim 1, wherein, The machine learning prediction model is a dynamic committee integration model.

8. The machine learning based reservoir shale content prediction method of claim 7, wherein, The method for constructing the dynamic committee integration model comprises the following steps: taking fuzzy C-means clustering as a dynamic integration gate network, dividing the data set into K clusters, and determining the optimal matching relationship between each expert model and the sub-data set by using the fuzzy C-means clustering method; allocating weights to the expert models in each cluster; for each sample in the data set, the membership degree of the sample in each cluster is calculated; according to the membership degree of the sample in each cluster and the model weight of the corresponding cluster, the membership degree of the sample and the model weight value of the corresponding cluster are multiplied, so as to obtain the weight allocation result of each model to each sample in each cluster, and the weight value allocation value of each model to a certain sample in all clusters is accumulated to obtain the dynamic weight of the expert model to the sample; the dynamic weights of the expert models to the sample are used to perform weighted integration and combination of the prediction results of the expert models, as the dynamic committee integration model.

9. The machine learning based reservoir shale content prediction method of claim 7 or 8, wherein, The committee machine prediction model is constructed by using four expert models of BP neural network, LightGBM, GRNN and Dilateformer.

10. The machine learning based reservoir shale content prediction method of claim 1, wherein, The machine learning prediction model is one of BP neural network, LightGBM, GRNN and Dilateformer.

Images