Well logging big data-based total organic carbon content machine learning prediction method and device

Abstract

The invention relates to a total organic carbon content machine learning prediction method and device based on logging big data, which integrate multiple algorithms such as evolution, swarm intelligence and physical inspiration through an algorithm pool, provide rich search strategy choices for subsequent hyper-parameter optimization, and prevent a single algorithm from falling into local optimum. And collaborative optimization of algorithm search efficiency and precision is realized through quantitative evaluation of convergence iteration times and a final fitness value. Meanwhile, the convergence iteration times and the final fitness value are normalized into comparable scores, the objectivity of algorithm screening is improved, and manual subjective judgment deviation is avoided. Through correlation coefficient calculation of an optimal solution trajectory sequence, an algorithm combination with large search path difference is screened out, the complementarity algorithm combination reduces the sensitivity to noise data through a diversified search strategy, high stability can still be kept in a small sample scene, and the problem that a traditional machine learning model is prone to overfitting is solved. The total organic carbon content prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the fields of big data and machine learning technology, and more specifically, to a machine learning method and apparatus for predicting total organic carbon content based on well logging big data. Background Technology

[0002] In the field of oil and gas exploration and development, total organic carbon (TOC) content is a core indicator for evaluating the hydrocarbon generation potential of source rocks, and its accurate determination is of vital strategic significance for the exploration and development of resources such as shale gas and tight oil. Traditional TOC determination methods mainly rely on laboratory analysis of core samples, directly determining organic carbon content through chemical methods such as combustion and wet oxidation. However, such methods have significant technical limitations: they require the collection of physical core samples, resulting in high sampling costs and sparse sampling points; laboratory analysis procedures are complex and time-consuming, making it difficult to meet the needs of large-scale rapid exploration; and in deep strata or complex geological conditions, sample collection is difficult, restricting the acquisition of spatial continuity of TOC data.

[0003] With the advancement of well logging technology, TOC prediction methods based on big data from well logging have gradually become a research hotspot. Well logging data is a continuous data stream of rock physical properties collected from wellbore shafts hundreds or even thousands of meters underground. The volume of well logging data is enormous; for example, a conventional well at a depth of 3000-5000 meters, using more than a dozen instruments, can generate raw data ranging from several gigabytes to tens of gigabytes. Well logging data is not only large in volume, but also represents a typical collection of multi-dimensional, high-value geophysical big data. Existing technical approaches are mainly divided into two categories: empirical formula methods and machine learning methods. Empirical formula methods mainly include the Schmoker density method and Passey's ΔlogR technique. The Schmoker density method establishes a prediction model based on the linear relationship between density logging data and TOC; Passey's ΔlogR technique utilizes the superposition difference between resistivity and porosity logging curves to construct a TOC prediction equation. These methods are computationally simple and easy to promote, but they are highly dependent on empirical coefficients specific to geological conditions. Their adaptability decreases significantly in areas with complex geological structures and drastic lithological changes, and the prediction error often exceeds the allowable range for engineering applications. Machine learning methods, by introducing algorithms such as support vector machines, random forests, and neural networks, can capture the nonlinear mapping relationship between well logging big data and TOC, thus improving the prediction accuracy compared to empirical formula methods.

[0004] However, current methods for predicting TOC content using machine learning models still have limitations, resulting in low accuracy in TOC content prediction. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a machine learning prediction method and apparatus for total organic carbon content based on well logging big data, so as to solve one or more of the above-mentioned problems in the prior art.

[0006] To achieve the above objectives, in one aspect, embodiments of this application provide a machine learning method for predicting total organic carbon content based on well logging big data, including the following steps:

[0007] Acquire big data from well logging and establish an algorithm pool;

[0008] Run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count, and final fitness value for each algorithm;

[0009] The comprehensive performance score of each algorithm is obtained based on the number of convergence iterations and the final fitness value.

[0010] The algorithm corresponding to the maximum value in the overall performance score is identified as the target algorithm;

[0011] Based on the trajectory sequence of each optimal solution, the complementarity score between any algorithm and the target algorithm is obtained, and the algorithm corresponding to the maximum value in the complementarity score is identified as the target algorithm, until the preset conditions are met.

[0012] The optimal hyperparameters are obtained by searching using various objective algorithms;

[0013] A basic model library was established, and the basic models in the basic model library were trained using the optimal hyperparameters to obtain multiple trained prediction models for total organic carbon content.

[0014] Based on various total organic carbon content prediction models and well logging big data, the prediction results of total organic carbon content were obtained.

[0015] In one embodiment, the step of obtaining the comprehensive performance score of each algorithm based on the number of convergence iterations and the optimal fitness value includes:

[0016] The convergence speed index is obtained by processing the number of convergence iterations.

[0017] Process the final fitness value to obtain the solution quality index;

[0018] A comprehensive performance score is obtained based on the convergence speed index and the solution quality index.

[0019] In one embodiment, the overall performance score is obtained based on the following formula:

[0020] ;

[0021] in, This is an indicator of convergence speed; To solve the quality indicators; The overall performance score; It is the largest proportion; It is the second largest proportion.

[0022] In one embodiment, the step further includes:

[0023] The reciprocal of the number of convergence iterations is determined as the convergence speed index, and the reciprocal of the final fitness value is determined as the solution quality index.

[0024] In one embodiment, the step of obtaining the complementarity score between any algorithm and the target algorithm based on the optimal solution trajectory sequence includes:

[0025] Calculate the correlation coefficient between the optimal solution trajectory sequences of any algorithm and the target algorithm;

[0026] The complementarity score is determined based on the correlation coefficient.

[0027] In one embodiment, the step of searching for optimal hyperparameters using various objective algorithms includes:

[0028] Initialize the random population for each objective algorithm;

[0029] A fitness function is constructed based on mean squared error, mean absolute error, and coefficient of determination.

[0030] Establish a public elite individual pool; the public elite individual pool is used to store high-quality candidate solutions obtained by each objective algorithm during the search process;

[0031] The target algorithms are iterated in parallel. During the iteration of any target algorithm, a preset number of elite individuals are randomly selected from the common elite individual pool, and the low-fitness individuals of the target algorithm are replaced with elite individuals.

[0032] Obtain the contribution of each target algorithm to the public elite individual pool at the current stage;

[0033] The computational resources for the next stage are adjusted based on the contribution level until the termination condition is met; computational resources include the number of iterations and / or the population size.

[0034] Calculate the fitness of individuals in the current public elite pool based on the fitness function;

[0035] The individual with the highest fitness is selected as the optimal hyperparameter.

[0036] In one embodiment, the step of obtaining the total organic carbon content prediction result based on various total organic carbon content prediction models and well logging big data includes:

[0037] By inputting the big data from well logging into various total organic carbon content prediction models, multiple current predicted values ​​are obtained;

[0038] Based on the normalized weights, the current predicted values ​​are summed in a weighted manner to obtain the current weighted fusion prediction result.

[0039] Based on the current predicted value and the current weighted prediction result, the current enhanced feature vector is obtained;

[0040] The current enhanced feature vector is input into the trained meta-learner to obtain the prediction result of total organic carbon content.

[0041] In one embodiment, the step further includes:

[0042] Obtain the coefficients of determination for each total organic carbon content prediction model on the validation set;

[0043] Based on the coefficient of determination, the normalized weights of any total organic carbon content prediction model are obtained;

[0044] Based on normalized weights, the original predicted values ​​output by each total organic carbon content prediction model are weighted and summed to obtain the initial weighted fusion prediction result.

[0045] Based on the initial weighted fusion prediction results and the original prediction values, an initial enhanced feature vector is obtained;

[0046] Using the initial enhanced feature vectors, a meta-learner is trained to obtain the trained meta-learner.

[0047] In one embodiment, the step further includes:

[0048] Interpretability analysis was performed on the total organic carbon content prediction results and well logging big data to obtain global feature importance ranking, feature dependency graph and single sample prediction interpretation;

[0049] Uncertainty quantification analysis was performed on the predicted total organic carbon content to obtain the corresponding uncertainty quantification results.

[0050] On the other hand, embodiments of the present invention provide a machine learning prediction device for total organic carbon content based on well logging big data, comprising:

[0051] The algorithm pool creation module is used to acquire big data from well logging and to create an algorithm pool.

[0052] The pre-run module is used to run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count and final fitness value of each algorithm;

[0053] The comprehensive performance evaluation module is used to obtain the comprehensive performance score of each algorithm based on the number of convergence iterations and the final fitness value.

[0054] The first target algorithm determination module is used to identify the algorithm corresponding to the maximum value in the comprehensive performance score as the target algorithm.

[0055] The second target algorithm determination module is used to obtain the complementarity score between any algorithm and the target algorithm based on the trajectory sequence of each optimal solution, and to identify the algorithm corresponding to the maximum value of the complementarity score as the target algorithm until the preset conditions are met.

[0056] The optimal hyperparameter acquisition module is used to search for and obtain the optimal hyperparameters using various objective algorithms.

[0057] The training module is used to establish a basic model library and to train the basic models in the basic model library using the optimal hyperparameters to obtain multiple trained total organic carbon content prediction models.

[0058] The prediction module is used to obtain the prediction results of total organic carbon content based on various total organic carbon content prediction models and well logging big data.

[0059] Compared with existing technologies, the machine learning prediction method for total organic carbon content based on well logging big data provided by this invention can achieve the following beneficial effects:

[0060] By integrating multiple algorithms, including evolutionary, swarm intelligence, and physics-inspired algorithms, a rich selection of search strategies is provided for subsequent hyperparameter optimization, preventing single algorithms from getting trapped in local optima. Quantitative evaluation of convergence iterations and final fitness values ​​achieves synergistic optimization of algorithm search efficiency and accuracy. Simultaneously, the convergence iteration count and final fitness value are normalized into comparable scores, improving the objectivity of algorithm selection and avoiding bias from subjective human judgment. By calculating the correlation coefficient of the optimal solution trajectory sequence, algorithm combinations with significant differences in search paths are selected. Complementary algorithm combinations reduce sensitivity to noisy data through diversified search strategies, maintaining high stability even in small sample scenarios and addressing the overfitting problem of traditional machine learning models. Multi-algorithm collaborative hyperparameter search achieves high accuracy and robustness in TOC prediction, improving noise resistance, reducing overfitting risk, and enhancing prediction accuracy compared to single machine learning models. Attached Figure Description

[0061] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic flowchart of a machine learning method for predicting total organic carbon content based on well logging big data in one embodiment;

[0064] Figure 2 This is a schematic flowchart illustrating the steps of obtaining the complementarity score between any algorithm and the target algorithm based on the trajectory sequences of each optimal solution in one embodiment.

[0065] Figure 3 This is a schematic flowchart illustrating the steps of searching for the optimal hyperparameters using various objective algorithms in one embodiment;

[0066] Figure 4 This is a schematic flowchart illustrating the steps for obtaining total organic carbon content prediction results based on various total organic carbon content prediction models and well logging big data in one embodiment.

[0067] Figure 5 This is a scatter plot showing the correlation between various logging parameters and TOC content in one embodiment;

[0068] Figure 6 This is a visualization of the dimensionality reduction of principal component analysis (PCA) in one embodiment;

[0069] Figure 7 This is an integrated model prediction accuracy verification graph in one embodiment, including a scatter plot of predicted and measured values ​​from the training and test sets;

[0070] Figure 8 This is a SHAP global feature importance analysis diagram in one embodiment, showing the ranking of the contributions of each logging parameter to TOC prediction;

[0071] Figure 9 This is a SHAP waterfall plot in one embodiment, showing the specific contribution of each feature of a single sample to the prediction result;

[0072] Figure 10 Here is a SHAP dependency plot in one embodiment, showing the nonlinear relationship between various logging parameters and TOC;

[0073] Figure 11 This is a TOC prediction result diagram for an actual well section in one embodiment, showing the comparison between the predicted TOC curve and the measured value within the depth range of 3085-3170 meters, as well as the uncertainty quantification results of the 95% confidence interval. Detailed Implementation

[0074] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0076] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustration and has no specific meaning in itself. Therefore, "module" and "part" may be used interchangeably.

[0077] It is understood that the term "connection" in the following embodiments should be understood as "electrical connection," "communication connection," etc., if the connected circuits, modules, units, etc., have electrical signal or data transmission with each other.

[0078] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising,” “including,” or “having,” etc., specify the presence of the stated feature, whole, step, operation, component, part, or combination thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof.

[0079] In one embodiment, such as Figure 1 As shown, a machine learning method for predicting total organic carbon content based on well logging big data is provided, including the following steps:

[0080] S110 acquires big data from well logging and establishes an algorithm pool;

[0081] The well logging big data includes any one or any combination of the following parameters: compensated neutrons, potassium content, bulk density, shallow resistivity, deep resistivity, uranium content, thorium content, acoustic transit time, photoelectric factor, and spontaneous potential. The algorithms in the algorithm pool can include biologically inspired algorithms (genetic algorithm, differential evolution algorithm, artificial bee colony algorithm, ant colony algorithm), animal behavior inspired algorithms (whale optimization algorithm, gray wolf optimization algorithm, Harris eagle optimization algorithm, marine predator algorithm, butterfly optimization algorithm), swarm intelligence algorithms (particle swarm optimization algorithm, firefly algorithm, bat algorithm, cuckoo search algorithm), physically inspired algorithms (simulated annealing algorithm, artificial plant optimization algorithm, gravity search algorithm, evolutionary multi-objective optimization algorithm), and mathematically inspired algorithms (sine and cosine algorithms, non-dominated sorting genetic algorithm II, teaching optimization algorithm), among others.

[0082] Specifically, raw well logging big data can be collected from oil and gas field databases, well logging companies, or public datasets. To eliminate quality issues in the acquisition process of the aforementioned well logging data, preprocessing of the well logging big data is necessary. In one embodiment, missing value handling uses the K-nearest neighbor imputation method, which calculates the Euclidean distance between missing samples and complete samples, and selects the k nearest neighbor samples for weighted average imputation. Outlier detection uses the 3σ criterion; when a well logging parameter value deviates from the mean by more than a preset standard deviation multiple, it is determined to be an outlier. Feature standardization uses the Z-score method or the min-max standardization method to convert each well logging parameter into a standard normal distribution. Dataset partitioning adopts a stratified sampling strategy, dividing the dataset into training and test sets in an appropriate proportion to ensure the uniformity of sample distribution across each TOC content interval.

[0083] S120: Run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count, and final fitness value for each algorithm;

[0084] The optimal solution trajectory sequence refers to the sequence of current optimal solutions (i.e., hyperparameter combinations) generated in each iteration during the iterative optimization process, reflecting the algorithm's search path and convergence trend. The convergence iteration count refers to the number of iterations required for the algorithm to reach the preset convergence condition (e.g., the fitness value change is less than a threshold), reflecting the algorithm's convergence speed. The final fitness value refers to the final value of the objective function (e.g., the reciprocal of the prediction error or the correlation coefficient) at the end of the algorithm's optimization, reflecting the model's performance under that hyperparameter combination.

[0085] Specifically, candidate algorithms are run on a standard test function set (Sphere, Rastrigin, Ackley, etc.). For each algorithm in the algorithm pool, using well logging big data as input and prediction error (e.g., RMSE) or correlation coefficient (e.g., R²) as the objective function, iterative optimization is performed. In each iteration, the current optimal solution (hyperparameter combination) and its corresponding fitness value are recorded, forming an optimal solution trajectory sequence. When the fitness value change is less than a preset threshold or the maximum number of iterations is reached, optimization stops, and the convergence iteration count and final fitness value are recorded. The fitness value can be calculated with reference to the fitness function.

[0086] S130, based on the number of convergence iterations and the final fitness value, the comprehensive performance score of each algorithm is obtained;

[0087] The overall performance score is a comprehensive evaluation index calculated based on the number of convergence iterations and the final fitness value. It can be processed by weighted summation or normalization to quantify the overall performance of the algorithm.

[0088] Specifically, the convergence iteration count is processed to obtain a convergence speed index; the final fitness value is processed to obtain a solution quality index; and a comprehensive performance score is obtained based on the convergence speed index and the solution quality index. In one embodiment, the method further includes the steps of determining the convergence speed index as the reciprocal of the convergence iteration count and the solution quality index as the reciprocal of the final fitness value. In one embodiment, the comprehensive performance score is obtained based on the following formula:

[0089] ;

[0090] in, This is an indicator of convergence speed; To solve the quality indicators; The overall performance score; As the first proportion, it can take the value 0.4; As the second proportion, it can take the value 0.6.

[0091] S140, the algorithm corresponding to the maximum value in the comprehensive performance score is identified as the target algorithm;

[0092] Specifically, the algorithm is sorted from highest to lowest based on its overall performance score, and the algorithm with the highest score, A1, is selected as the target algorithm.

[0093] S150: Based on the trajectory sequence of each optimal solution, obtain the complementarity score between any algorithm and the target algorithm, and identify the algorithm corresponding to the maximum value in the complementarity score as the target algorithm, until the preset conditions are met.

[0094] Specifically, the process of calculating the complementarity score is as follows: For each algorithm in the algorithm pool other than the target algorithm, calculate the correlation coefficient between its optimal solution trajectory sequence and the target algorithm's trajectory sequence. The complementarity score is calculated as 1 - the correlation coefficient. For two optimal solution trajectory sequences... Its correlation coefficient ;in, for covariance, for Standard deviation; for The standard deviation.

[0095] Sort the algorithms by complementarity score from highest to lowest, select algorithm A2 corresponding to the highest complementarity score, and include it in the target algorithm. Then repeat the above steps, that is, calculate the average complementarity between the remaining algorithms and {A_1, A_2}, and select the one with the highest score as A3. The preset condition, which is also the termination condition, is to select... Algorithms ( or marginal complementary gain .

[0096] S160, using various objective algorithms to search for the optimal hyperparameters;

[0097] Hyperparameters are parameters that need to be manually set before model training begins. They control the behavior of the learning algorithm, the complexity of the model, and the training speed. For a random forest model, the optimal hyperparameters might be the number of trees in the forest and the maximum depth of the trees. For a support vector machine, the optimal hyperparameters might be the penalty coefficient and the kernel function parameters. For a neural network model, the optimal hyperparameters might be the number of hidden layers, the number of neurons in each layer, and the learning rate. Optimal hyperparameters refer to a set of parameters found within a large search space. This set of parameters minimizes the objective function of the model when processing well logging data. In other words, this set of parameters minimizes the error, maximizes the accuracy, and maximizes the generalization ability when predicting TOC.

[0098] Specifically, a multi-algorithm collaborative search strategy can be adopted, using the target algorithm to search for the optimal combination in the hyperparameter space.

[0099] S170, establish a basic model library, and use the optimal hyperparameters to train the basic models in the basic model library to obtain multiple trained total organic carbon content prediction models.

[0100] The basic model library includes at least four groups: linear models, tree models, boosting models, support vector machines (SVMs), neural networks, and Gaussian process regression. In one embodiment, models are built using linear model-based algorithms, including ridge regression, Lasso regression, elastic network regression, and Bayesian ridge regression. Linear models offer good interpretability and computational efficiency. Specifically, models can be built using tree model-based algorithms, including random forests, extreme random trees, decision tree regression, and adaptive boosting decision trees. Tree models can capture nonlinear relationships and interactions between features. Models can also be built using gradient boosting algorithms, including extreme gradient boosting trees, classification gradient boosting, lightweight gradient boosting machines, and gradient boosting regression trees. Gradient boosting algorithms achieve excellent prediction accuracy by progressively optimizing residuals. Furthermore, models can be built using SVM-based algorithms, including support vector regression, linear support vector regression, and kernel ridge regression. SVMs can handle high-dimensional nonlinear problems through kernel functions. Finally, models can be built using neural network-based algorithms, including multilayer perceptrons, radial basis function neural networks, and extreme learning machines. Neural networks possess powerful nonlinear fitting capabilities. Models are built using algorithms based on Gaussian processes, including Gaussian process regression and Gaussian mixture model regression. Gaussian processes can quantify prediction uncertainty. Models are also built using algorithms based on ensemble learning, including Bagging regression, voting regression, and stacked regression. Ensemble methods improve overall performance by combining multiple weak learners. Models are also built using proximity algorithms, including K-nearest neighbor regression and locally weighted regression. Proximity algorithms predict based on sample similarity. By constructing a multi-level ensemble learning framework covering six major categories of basic models, the complementary advantages of different models are fully utilized, and a multi-index objective function is employed to balance prediction accuracy and model stability.

[0101] Specifically, traditional single-model training methods, while improving model performance, limit the model's generalization ability. Ensemble learning, on the other hand, possesses excellent model fusion characteristics. Its base model varies in diversity, and different types of models exhibit different predictive and generalization capabilities. It can better characterize data features, perform arbitrarily refined learning on complex data patterns, and maximize the complementarity of the strengths and weaknesses of each model type. This not only improves prediction accuracy and reduces prediction variance but also maintains model stability, achieving the best fit to the data.

[0102] To better explain the implementation process of ensemble learning, the base models are trained using optimal hyperparameter configurations, as shown in the following equation:

[0103] ;

[0104] in, This represents the base model after training. Represents the basic model architecture. This represents the optimal hyperparameter configuration obtained in step S102. This represents the training dataset.

[0105] S180, based on various total organic carbon content prediction models and well logging big data, yielded the total organic carbon content prediction results.

[0106] The predicted total organic carbon content includes both the TOC content and the confidence interval.

[0107] Specifically, model ensemble methods (such as weighted averaging and stacking ensemble) can be used to fuse the outputs of multiple prediction models, thereby improving prediction accuracy and robustness. Alternatively, hierarchical ensemble methods can be used to fuse prediction results; these methods consist of two levels: the first level is a weighted average ensemble based on validation set performance, and the second level is a stacking ensemble. These two levels are executed sequentially to form a progressive ensemble.

[0108] The aforementioned machine learning prediction method for total organic carbon (TOC) based on well logging big data integrates multiple algorithms, including evolutionary, swarm intelligence, and physics-inspired algorithms, through an algorithm pool. This provides a rich selection of search strategies for subsequent hyperparameter optimization, preventing a single algorithm from getting trapped in local optima. Quantitative evaluation of the convergence iteration count and final fitness value achieves synergistic optimization of algorithm search efficiency and accuracy. Simultaneously, normalizing the convergence iteration count and final fitness value into comparable scores enhances the objectivity of algorithm selection and avoids bias from subjective human judgment. By calculating the correlation coefficient of the optimal solution trajectory sequence, algorithm combinations with significant differences in search paths are selected. Complementary algorithm combinations reduce sensitivity to noisy data through diversified search strategies, maintaining high stability even in small sample scenarios and addressing the overfitting problem of traditional machine learning models. The collaborative search of hyperparameters by multiple algorithms achieves high accuracy and strong robustness in TOC prediction. Compared to a single machine learning model, it improves noise resistance, reduces the risk of overfitting, and enhances prediction accuracy.

[0109] In one embodiment, such as Figure 2 As shown, the steps for obtaining the complementarity score between any algorithm and the target algorithm based on the trajectory sequences of each optimal solution include:

[0110] S210, calculate the correlation coefficient between the optimal solution trajectory sequence of any algorithm and the target algorithm;

[0111] S220, determine the complementarity score based on the correlation coefficient.

[0112] Specifically, for each algorithm in the algorithm pool other than the target algorithm, calculate the correlation coefficient between its optimal solution trajectory sequence and the target algorithm's trajectory sequence. The complementarity score is calculated as 1 - correlation coefficient. For two optimal solution trajectory sequences... Its correlation coefficient ;in, for covariance, for Standard deviation; for The standard deviation.

[0113] In one embodiment, such as Figure 3 As shown, the steps for searching for the optimal hyperparameters using various objective algorithms include:

[0114] S310, initialize the random population for each objective algorithm respectively;

[0115] In the optimization algorithm, the set consists of multiple randomly initialized candidate solutions (hyperparameter combinations). Each individual represents a possible hyperparameter configuration.

[0116] Specifically, the Latin hypercube sampling (LHS) method can be used to initialize the population of each optimization algorithm to ensure that the initial solution is uniformly distributed in the parameter space.

[0117] S320, based on mean squared error, mean absolute error and coefficient of determination, constructs a fitness function;

[0118] Specifically, the model performance is comprehensively evaluated through a multi-index objective function, which is:

[0119] ;

[0120] in, , , , These are weighting coefficients, which can be adjusted according to actual application requirements. Mean squared error; Mean absolute error; The coefficient of determination.

[0121] S330, Establish a public elite individual pool; the public elite individual pool is used to store high-quality candidate solutions obtained by each objective algorithm during the search process;

[0122] Specifically, establish a public elite individual pool with a capacity of E. This pool is used to store high-quality candidate solutions discovered during the search process of each algorithm. The size E of the elite pool is adaptively set according to the number of algorithms N and the problem dimension D, and is calculated by the formula E=min(N×10, 50).

[0123] S340, iterate each target algorithm in parallel, and during the iteration of any target algorithm, randomly select a preset number of elite individuals from the common elite individual pool, and replace the low-fitness individuals of the target algorithm with elite individuals;

[0124] Specifically, the selection criterion for elite individuals adopts the fitness threshold method. When the fitness value of a candidate solution of a certain objective algorithm is better than the fitness value of the worst individual in the current elite pool, the candidate solution enters the elite pool and replaces the worst individual. The elite pool is maintained using a priority queue data structure to ensure that the time complexity of insertion and deletion operations is O(logE).

[0125] Furthermore, taking the target algorithms as examples, PSO (Particle Swarm Optimization), WOA (Whale Optimization Algorithm), GA (Genetic Algorithm), and GWO (Grey Wolf Optimizer), these algorithms are executed in parallel, every... A cross-algorithm information exchange is performed, in which The specific method of information exchange is as follows: each algorithm randomly selects k elite individuals (k=3~5) from the public elite pool, with probability... Elite individuals are introduced into the population to replace the worst-fitting individual. For the PSO algorithm, elite individuals directly replace the worst-fitting particle and reset its velocity; for the GA algorithm, elite individuals participate in the next round of crossover as external superior genes; for the WOA and GWO algorithms, elite individuals act as additional guiding targets, influencing the update of the search direction.

[0126] S350, obtain the contribution of each target algorithm to the public elite individual pool at the current stage;

[0127] S360: Adjust the computing resources for the next stage based on the contribution until the termination condition is met; computing resources include the number of iterations and / or the population size.

[0128] Specifically, the allocation of search resources is dynamically adjusted based on the real-time performance of each algorithm during the optimization process. The algorithm weight w_i is calculated based on its historical contribution, defined as the weighted sum of the number of elite individuals contributed to the elite pool by the algorithm in the last M iterations and the improvement of the global optimal solution. The weight update formula is as follows:

[0129] ;

[0130] Among them, C i Let ΔF be the elite contribution number of algorithm i.i The improvement in the optimal solution brought by algorithm i. The learning rate (ranging from 0.1 to 0.3) is used. Algorithms with higher weights can obtain more fitness function evaluations, achieving adaptive optimization of computational resources. The termination condition is reaching the maximum number of iterations or fitness convergence.

[0131] Furthermore, during the iteration process, the average distance index of the population can be used. Population diversity is measured, where d(x) is the Euclidean distance. When the diversity falls below the threshold θ_div = 0.1 × SearchRange, a diversity recovery mechanism is triggered: velocity reset and location perturbation are performed on PSO, the mutation probability is increased to 0.2 on GA, and some individuals are randomly reset to the search space for WOA and GWO.

[0132] S370, Calculate the fitness of individuals in the current public elite pool based on the fitness function;

[0133] S380 uses the individual with the highest fitness as the optimal hyperparameter.

[0134] Specifically, after the iteration is completed (i.e. the termination condition is met), the optimal hyperparameters are output based on the fitness value.

[0135] The above method obtains the optimal hyperparameters (as shown in Table 1) through multivariate heuristic collaborative optimization. The optimal hyperparameter configurations for each model significantly improve model performance. The collaborative optimization algorithm pool automatically searches and determines the optimal hyperparameter configurations for various algorithms, including linear model groups, tree model groups, gradient boosting groups, support vector machine groups, neural network groups, and Gaussian process groups, achieving an overall improvement in model performance.

[0136] Table 1

[0137]

[0138] In one embodiment, such as Figure 4 As shown, the steps for obtaining the total organic carbon content prediction results based on various total organic carbon content prediction models and well logging big data include:

[0139] S410 inputs the big data from well logging into various total organic carbon content prediction models to obtain multiple current prediction values;

[0140] S420, based on normalized weights, performs a weighted summation of each current predicted value to obtain the current weighted fusion prediction result;

[0141] Specifically, the integration weights can be determined based on the performance evaluation results through the following three methods:

[0142] (1) Weight allocation based on R² value: The weights of each model are proportional to their R² values, and the weight calculation formula is as follows: K is the total number of basic models, and this method gives higher weights to models with strong predictive power. The coefficient of determination for model m; Let be the determination coefficient of model j.

[0143] (2) Weight allocation based on comprehensive performance score: Taking into account multiple indicators, the comprehensive score of each model is first calculated. ,in , , Weighting coefficients for indicators (default) =0.5, =0.3, =0.2), and then calculate the weight based on the comprehensive score. ;

[0144] Weight allocation based on ranking: Models are ranked according to their R² values, and weights are allocated using an exponential decay function. ,in λ is the ranking of model m (the one with the highest R² is ranked 1), and λ is the decay coefficient (values ​​range from 0.3 to 0.5).

[0145] A weighted average of multiple predicted values, such as the fusion result. ,in . This represents the nth predicted value. These are the corresponding normalized weights.

[0146] S430, based on the current predicted value and the current weighted prediction result, obtain the current enhanced feature vector;

[0147] S440 inputs the current enhanced feature vector into the trained meta-learner to obtain the total organic carbon content prediction result.

[0148] Specifically, the weighted fusion result, together with the original predictions of the K base models, constitutes a K+1 dimensional feature vector (the current enhanced feature vector), which serves as the input to the meta-learner. The meta-learner employs linear regression, ridge regression, or gradient boosting algorithms to learn the optimal combination relationship between the weighted fusion result and the predictions of each base model, further improving prediction accuracy. This avoids the limitations of using weighted averaging alone, where the weights are fixed and cannot be adaptively adjusted, and also avoids the potential overfitting risk when using Stacking alone. A hierarchical progressive fusion strategy achieves the optimal combination of model prediction results. The first level calculates the R² normalized weights of each model based on the validation set performance and performs weighted averaging fusion. The second level inputs the weighted fusion result and the original predictions of each base model into the Stacking meta-learner for secondary optimization. The two levels are not mutually exclusive but rather sequentially progressive. The weighted averaging layer provides the initial fusion result and retains the diversity information of each model's predictions, while the Stacking layer further learns the optimal combination relationship to improve the final prediction accuracy.

[0149] Furthermore, in some cases, more than 10 heuristic optimization algorithms are executed in parallel. While cooperative mechanisms can improve optimization capabilities, this full search leads to a geometric increase in computational costs (CPU / memory / time) when dealing with large-scale datasets. In addition, many algorithms (such as GWO and WOA) exhibit high similarity in certain parameter ranges, resulting in significant computational redundancy. First, a real-time performance scoring model is established for each algorithm in the algorithm pool, as shown in the following formula:

[0150] ;

[0151] Where Si is the overall performance score of algorithm i; w1, w2, and w3 are the weights corresponding to each algorithm, which can be set according to the actual situation. Let $i$ be the improvement of the global optimal fitness value of algorithm $i$ within the most recent observation window (e.g., 50 iterations). Let represent the number of individuals derived from algorithm i in the public elite pool (Elite_Pool). E is the total capacity of the public elite pool. Eeff,i represents the computational efficiency of algorithm i, which is equal to the ratio of fitness improvement to computation time consumed. The steps described above for searching optimal hyperparameters using various target algorithms can be divided into two stages, executing different degrees of algorithm reduction. In the first stage, all target algorithms (10+ types) are run at full speed in parallel, and algorithms with the lowest overall performance scores (bottom 30%) and relatively parallel fitness curves are eliminated. In the second stage, every fixed step (e.g., 50 generations), only algorithms with scores above the median are retained. The population slots (computing power) released by the eliminated algorithms are allocated to the 2-3 core algorithms with the highest current scores (e.g., PSO or GWO), increasing their population density to improve search precision.

[0152] In one embodiment, the step further includes:

[0153] Obtain the coefficients of determination for each total organic carbon content prediction model on the validation set;

[0154] Based on the coefficient of determination, the normalized weights of any total organic carbon content prediction model are obtained;

[0155] Based on normalized weights, the original predicted values ​​output by each total organic carbon content prediction model are weighted and summed to obtain the initial weighted fusion prediction result.

[0156] Based on the initial weighted fusion prediction results and the original prediction values, an initial enhanced feature vector is obtained;

[0157] Using the initial enhanced feature vectors, a meta-learner is trained to obtain the trained meta-learner.

[0158] Specifically, this step involves training a meta-learner using a validation set, and its explanation can be found in the above-mentioned embodiments regarding obtaining the prediction results of total organic carbon content.

[0159] In one embodiment, the step further includes:

[0160] Interpretability analysis was performed on the total organic carbon content prediction results and well logging big data to obtain global feature importance ranking, feature dependency graph and single sample prediction interpretation;

[0161] Uncertainty quantification analysis was performed on the predicted total organic carbon content to obtain the corresponding uncertainty quantification results.

[0162] Specifically, interpretability analysis employs the SHAP method, using TreeExplainer for fast and accurate calculations for tree models, LinearExplainer for linear models, and KernelExplainer for other complex models, generating a global feature importance ranking, feature dependency graphs, and single-sample prediction explanations. For example... Figure 5 As shown, the correlation analysis between various logging parameters and TOC content indicates that U content and CNL have a strong positive correlation with TOC, providing a basis for subsequent characteristic importance analysis. Figure 6 As shown, the principal component analysis (PCA) dimensionality reduction visualization plot shows the distribution characteristics of samples with different TOC content ranges in two-dimensional space. High TOC samples are mainly distributed in the positive direction of the PC1 axis and are positively correlated with features such as U (uranium content) and CNL (compensating neutrons), while low TOC samples are mainly distributed in the negative direction of the PC1 axis.

[0163] like Figure 7As shown, the method described in this paper demonstrates excellent prediction accuracy on both the training and test sets for predicting TOC content, thus verifying the effectiveness of the method.

[0164] like Figure 8 , Figure 9 , Figure 10 As shown, SHAP analysis results indicate that U content and CNL are the most critical factors affecting TOC prediction, providing a scientific basis for geological interpretation. Interpretability analysis was used to calculate the contribution of each feature to the prediction results. These analysis results have the following practical effects on the final prediction: (1) Geological rationality verification of the prediction results: The TOC relationship of the features learned by the model is verified by the feature importance ranking. For example, if the U content, CNL and other parameters known to be related to organic matter enrichment are ranked high, it indicates that the prediction results are geologically rational and enhances the credibility of the prediction results; (2) Feature screening and model optimization: Redundant features that contribute little to the TOC prediction are identified according to the feature importance ranking, which provides a basis for subsequent model simplification and feature engineering; (3) Single sample prediction interpretation: The SHAP waterfall plot is used to explain how the predicted value of each sample is obtained by the sum of the contributions of each feature. When the predicted value is abnormal, the feature that caused the abnormality can be traced back; (4) Nonlinear relationship identification: The nonlinear response relationship between each logging parameter and TOC is revealed by the feature dependency relationship diagram, which provides a quantitative basis for understanding the TOC occurrence law; (5) Decision support: Geological engineers can use this to determine which logging parameters are most critical to the TOC prediction and guide the optimization of logging schemes.

[0165] A resampling method is used to quantify the uncertainty of the prediction results. Specifically, the Bootstrap resampling method is used to calculate the confidence interval of the predicted values. The number of resampling iterations is adaptively determined within the range of 200-1000 iterations based on the sample size and accuracy requirements. For each resampling iteration, the mean and standard deviation of the prediction results are calculated, and a 95% confidence interval is constructed. Residual analysis is also performed, including tests for residual normality, homoscedasticity, and independence.

[0166] like Figure 11 The diagram shows the integrated logging interpretation results for the actual production well section (3085-3170 meters). From left to right, the diagram displays the following: depth channel, porosity and density channel (CNL / DEN), uranium and acoustic channel (U / AC), thorium and deep resistivity channel (TH / RD), potassium and shallow resistivity channel (K / RS), and TOC prediction results channel.

[0167] In the TOC prediction results, red dots represent measured TOC data from core samples, blue solid lines represent the integrated prediction TOC curve of the method of this invention, and light blue shaded areas represent the 95% confidence intervals calculated by Bootstrap resampling.

[0168] from Figure 11 The verification results show that:

[0169] (1) Verification of prediction accuracy: The predicted TOC curve (blue solid line) and the measured data points (red dots) of the core showed a very high degree of agreement throughout the entire depth range. Especially in the high TOC enrichment section of 3145m-3165m, the model accurately captured the dramatic trend of TOC content change without any obvious smoothing effect or lag phenomenon.

[0170] (2) Consistency of logging response: The model successfully learned the nonlinear mapping relationship between logging parameters and TOC. For example, in the high TOC area, there are characteristics of significantly increased U content (orange solid line), decreased bulk density DEN (dashed line), and increased sonic transit time AC (red dashed line), which is completely consistent with the geological response law of shale gas reservoirs.

[0171] (3) Validity of uncertainty quantification: The 95% confidence interval, indicated by the light blue shading, reasonably covers the vast majority of measured data points. It is noteworthy that in areas with fewer measured points or greater data fluctuations, the confidence interval widens appropriately, objectively reflecting the level of uncertainty in the prediction; while in areas with dense data and obvious patterns, the confidence interval narrows, indicating a higher level of model confidence. This interval estimation provides an important risk reference for practical engineering decisions, an advantage not possessed by traditional point prediction methods.

[0172] In summary, Figure 11 Actual well section verification fully demonstrates the effectiveness, robustness, and reliability of the method of this invention under complex geological conditions. The total organic carbon (TOC) prediction method based on multivariate heuristic collaborative optimization and ensemble learning in any of the above embodiments establishes a high-quality dataset through intelligent preprocessing, constructs a diverse model library using a multi-level ensemble learning framework, and obtains the optimal hyperparameter configuration through multivariate heuristic collaborative optimization. Furthermore, the prediction results of multiple optimization models are fused using an ensemble method, feature contribution explanations are provided through SHAP interpretability analysis, and uncertainty is quantified using the Bootstrap resampling method to obtain the final TOC prediction result. Based on this, the accuracy, efficiency, robustness, and reliability of TOC prediction are significantly improved. It maintains high performance even under small sample sizes, noisy data, and complex geological conditions, and is applicable to various reservoir types such as shale gas and tight oil, demonstrating good engineering practicality and broad application prospects.

[0173] In one implementation case, a machine learning prediction device for total organic carbon content based on well logging big data is provided, including:

[0174] The algorithm pool creation module is used to acquire big data from well logging and to create an algorithm pool.

[0175] The pre-run module is used to run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count and final fitness value of each algorithm;

[0176] The comprehensive performance evaluation module is used to obtain the comprehensive performance score of each algorithm based on the number of convergence iterations and the final fitness value.

[0177] The first target algorithm determination module is used to identify the algorithm corresponding to the maximum value in the comprehensive performance score as the target algorithm.

[0178] The second target algorithm determination module is used to obtain the complementarity score between any algorithm and the target algorithm based on the trajectory sequence of each optimal solution, and to identify the algorithm corresponding to the maximum value of the complementarity score as the target algorithm until the preset conditions are met.

[0179] The optimal hyperparameter acquisition module is used to search for and obtain the optimal hyperparameters using various objective algorithms.

[0180] The training module is used to establish a basic model library and to train the basic models in the basic model library using the optimal hyperparameters to obtain multiple trained total organic carbon content prediction models.

[0181] The prediction module is used to obtain the prediction results of total organic carbon content based on various total organic carbon content prediction models and well logging big data.

[0182] Specific limitations regarding the machine learning prediction device for total organic carbon content based on well logging big data can be found in the above-mentioned limitations on the machine learning prediction method for total organic carbon content based on well logging big data, and will not be repeated here. Each module in the aforementioned machine learning prediction device for total organic carbon content based on well logging big data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0183] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0184] Acquire big data from well logging and establish an algorithm pool;

[0185] Run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count, and final fitness value for each algorithm;

[0186] The comprehensive performance score of each algorithm is obtained based on the number of convergence iterations and the final fitness value.

[0187] The algorithm corresponding to the maximum value in the overall performance score is identified as the target algorithm;

[0188] Based on the trajectory sequence of each optimal solution, the complementarity score between any algorithm and the target algorithm is obtained, and the algorithm corresponding to the maximum value in the complementarity score is identified as the target algorithm, until the preset conditions are met.

[0189] The optimal hyperparameters are obtained by searching using various objective algorithms;

[0190] A basic model library was established, and the basic models in the basic model library were trained using the optimal hyperparameters to obtain multiple trained prediction models for total organic carbon content.

[0191] Based on various total organic carbon content prediction models and well logging big data, the prediction results of total organic carbon content were obtained.

[0192] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the following steps:

[0193] Acquire big data from well logging and establish an algorithm pool;

[0194] Run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count, and final fitness value for each algorithm;

[0195] The comprehensive performance score of each algorithm is obtained based on the number of convergence iterations and the final fitness value.

[0196] The algorithm corresponding to the maximum value in the overall performance score is identified as the target algorithm;

[0197] Based on the trajectory sequence of each optimal solution, the complementarity score between any algorithm and the target algorithm is obtained, and the algorithm corresponding to the maximum value in the complementarity score is identified as the target algorithm, until the preset conditions are met.

[0198] The optimal hyperparameters are obtained by searching using various objective algorithms;

[0199] A basic model library was established, and the basic models in the basic model library were trained using the optimal hyperparameters to obtain multiple trained prediction models for total organic carbon content.

[0200] Based on various total organic carbon content prediction models and well logging big data, the prediction results of total organic carbon content were obtained.

[0201] In specific implementation, the embodiments of this application can be referred to the above embodiments and have corresponding technical effects.

[0202] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0203] For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or external to the processor.

[0204] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0205] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0206] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0207] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0208] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0209] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0210] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A machine learning method for predicting total organic carbon content based on well logging big data, characterized in that, include: Acquire big data from well logging and establish an algorithm pool; Run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration count, and final fitness value for each algorithm; Based on the number of convergence iterations and the final fitness value, the comprehensive performance score of each algorithm is obtained; The algorithm corresponding to the maximum value in the comprehensive performance score is identified as the target algorithm; Based on the optimal solution trajectory sequence, the complementarity score between any of the algorithms and the target algorithm is obtained, and the algorithm corresponding to the maximum value of the complementarity score is identified as the target algorithm, until the preset conditions are met; The optimal hyperparameters are obtained by searching using the aforementioned target algorithms; A basic model library is established, and the basic models in the basic model library are trained using the optimal hyperparameters to obtain multiple trained total organic carbon content prediction models. Based on the total organic carbon content prediction models and the well logging big data, the total organic carbon content prediction results are obtained.

2. The machine learning prediction method for total organic carbon content based on well logging big data according to claim 1, characterized in that, The step of obtaining the comprehensive performance score of each algorithm based on the convergence iteration number and the optimal fitness value includes: The convergence iteration count is processed to obtain the convergence speed index; Process the final fitness value to obtain the solution quality index; The comprehensive performance score is obtained based on the convergence speed index and the solution quality index.

3. The machine learning prediction method for total organic carbon content based on well logging big data according to claim 2, characterized in that, The overall performance score is obtained based on the following formula: ； in, The convergence speed index is mentioned above; The quality index of the solution; The overall performance score; It is the largest proportion; It is the second largest proportion.

4. The machine learning prediction method for total organic carbon content based on well logging big data according to claim 2, characterized in that, It also includes the following steps: The reciprocal of the number of convergence iterations is determined as the convergence speed index, and the reciprocal of the final fitness value is determined as the solution quality index.

5. The machine learning method for predicting total organic carbon content based on well logging big data according to claim 1, characterized in that, The step of obtaining the complementarity score between any of the aforementioned algorithms and the target algorithm based on the optimal solution trajectory sequences includes: Calculate the correlation coefficient between the optimal solution trajectory sequence of any of the algorithms and the target algorithm; The complementarity score is determined based on the correlation coefficient.

6. The machine learning method for predicting total organic carbon content based on well logging big data according to claim 1, characterized in that, The steps for searching for optimal hyperparameters using the aforementioned target algorithms include: Initialize the random population for each of the target algorithms described above; A fitness function is constructed based on mean squared error, mean absolute error, and coefficient of determination. A public elite individual pool is established; wherein, the public elite individual pool is used to store high-quality candidate solutions obtained by each of the target algorithms during the search process; The target algorithms are iterated in parallel, and during the iteration of any target algorithm, a preset number of elite individuals are randomly selected from the public elite individual pool, and the low-fitness individuals of the target algorithm are replaced with the elite individuals. Obtain the contribution of each target algorithm to the public elite individual pool at the current stage; The computing resources for the next stage are adjusted based on the contribution level until the termination condition is met; the computing resources include the number of iterations and / or the population size. Calculate the fitness of individuals in the current public elite pool based on the fitness function. The individual with the highest fitness is taken as the optimal hyperparameter.

7. The machine learning method for predicting total organic carbon content based on well logging big data according to claim 1, characterized in that, The steps for obtaining the total organic carbon content prediction results based on the aforementioned total organic carbon content prediction models and the aforementioned well logging big data include: The well logging big data is input into each of the total organic carbon content prediction models to obtain multiple current prediction values; Based on the normalized weights, the current predicted values ​​are summed in a weighted manner to obtain the current weighted fusion prediction result. Based on the current predicted value and the current weighted prediction result, the current enhanced feature vector is obtained; The current enhanced feature vector is input into the trained meta-learner to obtain the predicted result of the total organic carbon content.

8. The machine learning prediction method for total organic carbon content based on well logging big data according to claim 7, characterized in that, It also includes the following steps: Obtain the coefficient of determination of each total organic carbon content prediction model on the validation set; Based on the determination coefficient, the normalized weights of any of the total organic carbon content prediction models are obtained; Based on the normalized weights, the original predicted values ​​output by each of the total organic carbon content prediction models are weighted and summed to obtain the initial weighted fusion prediction result. Based on the initial weighted fusion prediction result and the original prediction value, an initial enhanced feature vector is obtained; Using the initial enhanced feature vector, a meta-learner is trained to obtain the trained meta-learner.

9. The machine learning prediction method for total organic carbon content based on well logging big data according to claim 7, characterized in that, It also includes the following steps: Interpretability analysis was performed on the total organic carbon content prediction results and the well logging big data to obtain global feature importance ranking, feature dependency graph and single sample prediction interpretation; Uncertainty quantification analysis was performed on the predicted total organic carbon content to obtain the corresponding uncertainty quantification results.

10. A machine learning prediction device for total organic carbon content based on well logging big data, characterized in that, include: The algorithm pool creation module is used to acquire big data from well logging and to create an algorithm pool. The pre-run module is used to run the algorithms in the algorithm pool to obtain the optimal solution trajectory sequence, convergence iteration number and final fitness value of each algorithm; The comprehensive performance evaluation module is used to obtain the comprehensive performance score of each algorithm based on the number of convergence iterations and the final fitness value. The first target algorithm determination module is used to identify the algorithm corresponding to the maximum value in the comprehensive performance score as the target algorithm; The second target algorithm determination module is used to obtain the complementarity score between any of the algorithms and the target algorithm based on the optimal solution trajectory sequence, and to identify the algorithm corresponding to the maximum value of the complementarity score as the target algorithm until a preset condition is met. The optimal hyperparameter acquisition module is used to search for the optimal hyperparameters using the aforementioned target algorithms. The training module is used to establish a basic model library and to train the basic models in the basic model library using the optimal hyperparameters to obtain multiple trained total organic carbon content prediction models. The prediction module is used to obtain the prediction result of total organic carbon content based on the total organic carbon content prediction models and the well logging big data.

Images