Shale oil reservoir parameter calculation method and system, electronic device and storage medium

Abstract

This disclosure provides a method, system, electronic device, and storage medium for calculating shale oil reservoir parameters. The method includes acquiring datasets of engineering and geological parameters and preprocessing the datasets; defining multiple basic nonlinear regression models based on the preprocessed datasets and performing hyperparameter optimization on these models; selecting the hyperparameter-optimized basic nonlinear regression models and obtaining those that are not overfitted; performing dynamic weighted fusion based on the non-overfitted models; training the fused models to obtain a target nonlinear regression model; and obtaining shale oil reservoir parameters based on the target nonlinear regression model. The nonlinear regression model established in this way can effectively handle large and complex shale oil datasets and calculate shale oil reservoir characterization parameters, improving prediction accuracy and reliability, and meeting the demand for high-precision parameter calculation in shale oil exploration and development.

Description

Technical Field

[0001] This disclosure pertains to the field of oil and gas field exploration, and particularly relates to methods, systems, electronic equipment, and storage media for calculating shale oil reservoir parameters. Background Technology

[0002] Shale oil reservoirs refer to the pores and fractures within shale formations that contain shale oil. These pores are typically classified as intergranular pores, intercrystalline pores, and organic matter pores, characterized by low porosity and permeability. Accurately assessing the geological characteristics of shale oil (such as total organic carbon content, free hydrocarbons, vitrinite reflectance, porosity, and permeability) is fundamental to studying source rocks, reservoir characteristics, and is crucial for evaluating resource potential, reservoir quality, and development feasibility.

[0003] Currently, traditional empirical methods or multiple linear regression analysis are commonly used to calculate the geological characteristic parameters of shale oil. However, due to the large and complex volume of drilling and logging data, conventional methods struggle to effectively handle the inherent nonlinear relationships within the data, thus affecting the accuracy and reliability of shale oil geological characteristic parameter predictions. Summary of the Invention

[0004] To address the aforementioned issues, this disclosure provides a method, system, electronic equipment, and storage medium for calculating shale oil reservoir parameters. The adaptive regression model selection method and system using dynamic weighted fusion of multiple indicators are suitable for scenarios involving complex nonlinear data relationships and can improve the prediction accuracy and reliability of shale oil geological characteristic parameters.

[0005] The first aspect of this disclosure provides a method for calculating shale oil reservoir parameters. The method includes: acquiring a dataset of engineering parameters and geological parameters; preprocessing the dataset; defining multiple basic nonlinear regression models based on the preprocessed dataset; optimizing the hyperparameters of the multiple basic nonlinear regression models; screening the optimized basic nonlinear regression models and obtaining a basic nonlinear regression model that is not overfitted; performing dynamic weighted fusion based on the non-overfitted basic nonlinear regression model; training the fused basic nonlinear regression model to obtain a target nonlinear regression model; and obtaining shale oil reservoir parameters based on the target nonlinear regression model.

[0006] This setup, by considering nonlinear relationships in the preprocessed dataset, defines multiple basic nonlinear regression models. This allows for the optimization of hyperparameters in these models to obtain the optimal parameters. Furthermore, it facilitates fitting the optimal parameters from these models to identify those that are not overfitting. The fused model is then trained to obtain the target nonlinear regression model. This approach enables dynamic adjustment of model weights and fusion strategies, resulting in a high-performance, highly generalizable, and interpretable nonlinear regression model. The established nonlinear regression model is used to calculate reservoir characterization parameters, effectively handling large and complex shale oil datasets, accurately calculating geological characteristic parameters of shale oil, improving prediction accuracy and reliability, and meeting the high-precision parameter calculation requirements of shale oil exploration and development.

[0007] In some embodiments, the preprocessing of the dataset for obtaining engineering parameters and geological parameters includes: the dataset includes a training set of the engineering parameters and the geological parameters, and a test set of the engineering parameters and the geological parameters; the ratio of the training set to the test set is one of 6:4, 7:3, 8:2 or 9:1.

[0008] This setting helps handle missing values ​​and outliers, ensuring data quality.

[0009] In some embodiments, defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to these models includes: the basic nonlinear regression models are defined based on kernel functions, and the kernel functions include:

[0010] Given two samples x i and x j The kernel function K is defined as follows:

[0011] K(x i ,x j )=φ(x i ) T φ(x j )

[0012] Where φ is a nonlinear mapping function.

[0013] This setup, by defining kernel functions and calculating inner products, avoids explicitly calculating mapped features, reducing computational complexity. It allows for linear regression or classification in high-dimensional spaces, thereby improving the performance of nonlinear regression models.

[0014] In some embodiments, the various basic nonlinear regression models defined based on kernel functions include: kernel ridge regression, support vector regression, and Bayesian nonlinear regression.

[0015] This setup helps to prevent overfitting through kernel ridge regression regularization, improve model robustness by maximizing the margin through support vector regression, and provide uncertainty estimation and flexible nonlinear modeling capabilities through Bayesian nonlinear regression, thereby aiding in the solution of nonlinear problems.

[0016] In some embodiments, defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to the multiple basic nonlinear regression models includes: defining a tree model of the basic nonlinear regression models, wherein the tree model includes:

[0017] Divide the data, and at each node, select a feature x. i And the corresponding threshold t, divide the dataset into two parts, satisfying x i ≤t and x i >t;

[0018] Selecting the optimal split point involves optimizing the metric to choose the best features and split point: Among them, y i It is the actual value. It is a predicted value;

[0019] Recursively partition the tree, repeating the above steps on each child node until a stopping condition is met, such as when the number of samples in a node is less than a preset value or when the maximum tree depth is reached.

[0020] The prediction is the average of the target values ​​of the samples in the leaf nodes.

[0021] This setup constructs a decision tree by recursively partitioning the data space, thereby creating a set of rules for predicting new data, which is easy to interpret and handle complex nonlinear relationships.

[0022] In some embodiments, the various basic nonlinear regression models defined based on tree models include random forest regression and gradient boosting tree.

[0023] This setup reduces overfitting by averaging the results of multiple decision trees using a random forest, and improves the accuracy of the nonlinear regression model by gradually reducing residuals using a gradient boosting tree.

[0024] In some embodiments, defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to the multiple basic nonlinear regression models includes: defining the basic nonlinear regression models based on neural networks; the models defined by neural networks include multilayer perceptron regression.

[0025] With this setup, through multi-layered learning and weight adjustment, neural networks can reduce the risk of overfitting and improve model performance on new data.

[0026] In some embodiments, the multilayer perceptron regression includes:

[0027] The input layer accepts an n-dimensional input vector x = [x1, x2, ... xn]. n Each node represents a feature in the input data; a multilayer perceptron contains at least one hidden layer. Each neuron in the hidden layer calculates a weighted sum based on the input and weights, and applies an activation function for a non-linear transformation. The output of each neuron is: Where σ is the activation function, and These are the weights and biases; the output layer ultimately outputs the prediction result, providing a continuous value. in and b (2) These represent the weights and biases of the output layer, respectively, and m is the number of neurons in the hidden layer; during training, the predicted values ​​are calculated through forward propagation. Then, the prediction error is calculated based on the loss function; next, the gradient of the error with respect to each weight and bias is calculated using the backpropagation algorithm, and the weights and biases are updated using gradient descent. Where η is the learning rate, L is the loss function, and l represents the network layer.

[0028] This setup utilizes the nonlinear transformations of the multilayer perceptron regression layers to perform complex pattern learning and regression prediction on the input data, exhibiting strong nonlinear modeling capabilities and aiding in the definition of nonlinear regression models.

[0029] In some embodiments, defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters based on these models includes: locating the search range of hyperparameters for each basic nonlinear regression model; constructing a hyperparameter grid based on the search range; finding hyperparameter combinations for each basic nonlinear regression model based on the hyperparameter grid; evaluating the hyperparameter combinations for each basic nonlinear regression model to determine the target hyperparameter combination for each model.

[0030] This setup, by defining the search range of hyperparameters for each basic nonlinear regression model, helps to find the optimal combination of hyperparameters for each basic nonlinear regression model using grid search combined with cross-validation. Then, by evaluating the hyperparameter combinations, a target hyperparameter combination is selected, which can improve the performance and computational efficiency of the nonlinear regression model.

[0031] In some embodiments, the method for finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes grid search, random search, and Bayesian optimization methods.

[0032] This configuration allows for flexible selection based on different models and data characteristics, helping to achieve optimal performance with fewer computing resources.

[0033] In some embodiments, finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: finding the hyperparameter combination of each basic nonlinear regression model based on the grid search combined with cross-validation.

[0034] This setting allows for the combination of parameters one by one in the preset hyperparameter grid, enabling a comprehensive evaluation of model performance, and is suitable for situations with a small number of parameters.

[0035] In some embodiments, finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: training on a randomly selected hyperparameter combination in the hyperparameter grid based on the random search to find the hyperparameter combination of each basic nonlinear regression model.

[0036] This setup allows for the random selection of hyperparameter combinations for training, making it more time-efficient and suitable for high-dimensional hyperparameter spaces.

[0037] In some embodiments, finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: constructing a surrogate model based on the Bayesian optimization method and obtaining the iterative selection of the function to find the hyperparameter combination of each basic nonlinear regression model.

[0038] This setup allows for a more efficient use of a limited number of experiments to find the global optimum, thus enabling a more efficient search for the hyperparameter combination of the basic nonlinear regression model.

[0039] In some embodiments, evaluating the hyperparameter combinations of each basic nonlinear regression model to determine the target hyperparameter combination of each basic nonlinear regression model includes: evaluating the hyperparameter combinations of each basic nonlinear regression model using negative mean square error to determine the target hyperparameter combination.

[0040] This setting, using the negative mean square error to reflect the squared mean of the difference between the predicted and actual values, helps to make the target hyperparameter combination closer to the real data.

[0041] In some embodiments, the process of screening the hyperparameter-optimized basic nonlinear regression models and obtaining basic nonlinear regression models that are not overfitting includes: calculating multiple performance metrics for each basic nonlinear regression model; setting a difference threshold between the training set and cross-validation based on each performance metric; and determining and screening basic nonlinear regression models that are not overfitting based on the difference threshold between the training set and cross-validation.

[0042] This setup facilitates the judgment of the difference threshold of the basic nonlinear regression model under various performance indicators. Overfitting basic nonlinear regression models are discarded, while nonfitting basic nonlinear regression models are retained, thus making it easier to screen for nonfitting basic nonlinear regression models.

[0043] In some embodiments, the various performance metrics include: mean square error, mean absolute error, coefficient of determination, root mean square error, and mean absolute percentage error.

[0044] This setting helps determine whether the basic nonlinear regression model is overfitting.

[0045] In some embodiments, the screening of hyperparameter-optimized basic nonlinear regression models and obtaining non-overfitted basic nonlinear regression models includes: performing a diversity assessment on the non-overfitted basic nonlinear regression models and obtaining the correlation of the model residuals; and ranking the non-overfitted basic nonlinear regression models according to the correlation of the model residuals.

[0046] This setting allows for the prioritization of basic nonlinear regression models with low residual correlation based on the correlation of the residuals in the calculation models, which helps to increase the effectiveness of the fusion of basic nonlinear regression models.

[0047] In some embodiments, the dynamic weighted fusion based on the non-overfitted basic nonlinear regression model includes: dynamically calculating the weights of the selected basic nonlinear regression models based on the multiple performance indicators; and adaptively fusion of the selected basic nonlinear regression models based on the results of the dynamic weight calculation.

[0048] This setup, by combining the advantages of multiple models, avoids the limitations that may arise from a single model and reduces the risk of decision-making errors caused by a single model. At the same time, by fusing data, multiple aspects can be captured, enhancing overall performance and thus improving the accuracy of prediction results.

[0049] In some embodiments, the strategy for adaptively fusing the selected basic nonlinear regression models based on the dynamic weight calculation results includes weighted average, stacked regression, and voting method.

[0050] With this setup, the above strategy allows the fusion method to balance the performance of each model and reduce the negative impact caused by the failure of a single model when a single model has a high error rate.

[0051] In some embodiments, training and learning multiple fused basic nonlinear regression models to obtain a target nonlinear regression model, and obtaining shale oil reservoir parameters based on the target nonlinear regression model, includes: weighting the selected basic nonlinear regression models based on the stacked regression to form a base model; selecting one of multiple basic nonlinear regression models as a meta-learner based on the base model; training the stacked model based on the selected meta-learner to obtain the target nonlinear regression model; and obtaining shale oil reservoir parameters based on the target nonlinear regression model.

[0052] This setup serves two purposes: firstly, it facilitates the understanding and integration of the complex relationships between various basic nonlinear regression models after training them through a meta-learning period, thereby deriving the common patterns and optimal decision-making strategies of the basic nonlinear regression models; secondly, it enables the making of comprehensive judgments based on the prediction distributions of all models, addressing the uncertainties of the basic models, and thus optimizing the predictive capabilities of the target nonlinear regression model.

[0053] In some embodiments, the method further includes: evaluating the target nonlinear regression model after training; and performing online learning and model updating based on the evaluated target nonlinear regression model.

[0054] This setting facilitates the improvement and updating of the target nonlinear regression model, thereby enhancing its applicability.

[0055] In some embodiments, the evaluation based on the trained target nonlinear regression model includes: calculating multiple performance metrics of the fused target nonlinear regression model on a test set based on the base models; comparing the target nonlinear regression model with a single model among multiple base nonlinear regression models based on the calculation results; and verifying the superiority of the fused target nonlinear regression model based on the comparison results.

[0056] This setup helps to verify the superiority of the nonlinear regression model after fusion learning, and to identify a superior nonlinear regression model, thus validating the nonlinear regression model and aiding in the calculation of the characterization parameters of the oil reservoir.

[0057] A second aspect of this disclosure provides a shale oil reservoir parameter calculation system, the system comprising:

[0058] The system includes a data preprocessing module for acquiring and preprocessing datasets of engineering and geological parameters; a model definition and hyperparameter module for defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing the hyperparameters of these models; a model selection module for selecting the optimized basic nonlinear regression models and obtaining models that are not overfitted; a dynamic weighted fusion module for dynamically weighting and fusing the selected basic nonlinear regression models; and a training module for training the fused basic nonlinear regression models to obtain a target nonlinear regression model and acquiring shale oil reservoir parameters based on this target model.

[0059] The technical effects of any possible implementation of the second aspect can be found in the technical effects of the first aspect mentioned above, and will not be repeated here.

[0060] A third aspect of this disclosure provides an electronic device, the electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the shale oil reservoir parameter calculation method as described in the first aspect.

[0061] The technical effects of any possible implementation of the third aspect can be found in the technical effects of the first aspect mentioned above, and will not be repeated here.

[0062] The fourth aspect of this disclosure provides a computer storage medium storing a computer program that, when executed by a processor, implements the shale oil reservoir parameter calculation method as described in the first aspect.

[0063] The technical effects of any possible implementation of the fourth aspect can be found in the technical effects of the first aspect mentioned above, and will not be repeated here.

[0064] Compared with the prior art, this disclosure has the following advantages:

[0065] 1. For the calculation and prediction of geological characteristic parameters of shale oil, this paper comprehensively considers the nonlinear relationship of data, multiple performance indicators and model diversity, performs hyperparameter tuning, dynamically adjusts model weights and fusion strategies, integrates uncertainty estimation, provides online learning and model update mechanisms, and constructs a regression model with superior performance, strong generalization ability and high interpretability.

[0066] 2. It can effectively process large and complex shale oil datasets, improve the accuracy and reliability of predictions, provide a solid scientific basis for shale oil exploration and development, and has high practical value and promotion significance.

[0067] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 A flowchart illustrating the shale oil reservoir parameter calculation method provided in this embodiment of the disclosure;

[0070] Figure 2 A flowchart of the model definition and hyperparameter optimization method provided in the embodiments of this disclosure;

[0071] Figure 3 A flowchart illustrating the model evaluation and screening method provided in this disclosure embodiment;

[0072] Figure 4 A flowchart of a dynamic weighted fusion method provided in this embodiment of the disclosure;

[0073] Figure 5 A schematic diagram of porosity calculated based on a dynamic weighted nonlinear regression model, provided as an embodiment of this disclosure;

[0074] Figure 6 A block diagram of a reservoir characterization parameter calculation system provided in this embodiment of the disclosure;

[0075] Figure 7 This is a structural block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0077] See Figure 1 , Figure 1This is a flowchart of a shale oil reservoir parameter calculation method provided in this embodiment. Based on a dynamically weighted nonlinear regression model, it calculates reservoir characterization parameters. The aim is to utilize multiple nonlinear regression models, employing multi-index evaluation and dynamic weighted fusion strategies, combined with hyperparameter tuning and model diversity evaluation, to construct a high-performance regression model. This model can effectively handle large and complex shale oil datasets, accurately calculate geological characteristic parameters of shale oil, improve prediction accuracy and reliability, and meet the high-precision parameter calculation requirements of shale oil exploration and development. The method includes the following steps S100–S600:

[0078] S100. Obtain the dataset of engineering parameters and geological parameters, and preprocess the dataset.

[0079] In this embodiment, engineering parameters include drilling, logging, and well logging data, while geological parameters include porosity, a geological characteristic parameter of shale oil. Porosity is one of the geological characteristic parameters of shale oil, and it is calculated using logging data of natural gamma ray (GR), resistivity (RLLD), acoustic travel time (AC), calibrated neutron (CNL), and density (DEN).

[0080] It should be noted that the data read includes datasets of feature variables X and target variables y. Feature variable X consists of engineering parameters and other data used to explain and predict changes in the dependent variable; target variable y is the shale oil geological characteristic parameter, porosity. The read data will then be cleaned. Cleaning can be done programmatically using Python or MATLAB to handle missing values ​​and outliers, ensuring data quality.

[0081] It is understood that the dataset includes a training set and a test set of engineering and geological parameters; the dataset is divided into training and test sets according to one of the following ratios: 6:4, 7:3, 8:2, or 9:1, such as training set (X). train ,y train ), test set (X test ,y test The ratio of training set to test set is 7:3; the feature variables are standardized to have a mean of 0 and a variance of 1.

[0082] Feature variables

[0083] Where μ is the characteristic mean and σ is the characteristic standard deviation.

[0084] S200 defines various basic nonlinear regression models based on the preprocessed dataset and optimizes the hyperparameters of these models.

[0085] In this embodiment, hyperparameter optimization is performed based on multiple preprocessed basic nonlinear regression models to obtain the optimal hyperparameter combination for each basic nonlinear regression model.

[0086] In some embodiments, see Figure 2 , Figure 2 This is a flowchart of the model definition and hyperparameter optimization method provided in the embodiments of this disclosure. S200 includes S201 to S202:

[0087] S201. Define a variety of basic nonlinear regression models.

[0088] In this embodiment, various basic nonlinear regression models include, but are not limited to, kernel ridge regression, support vector regression and Bayesian nonlinear regression based on kernel function definition, random forest regression and gradient boosting tree based on tree model definition, and multilayer perceptron regression based on neural network definition.

[0089] For example, the basic nonlinear regression model is a regression analysis defined based on a kernel function. The kernel function method is a technique that achieves linear separability when the original space is linearly inseparable by mapping input data to a high-dimensional feature space. Its core idea is to reduce computational complexity by using inner product calculations to avoid explicitly calculating the mapped features. Let there be two samples x. i and x j The kernel function K is defined as:

[0090] K(x i ,x j )=φ(x i ) T φ(x j (2)

[0091] Here, φ is the mapping function, which is typically non-linear. Common kernel functions include:

[0092] Linear kernel: K(x) i ,x j )=x i T x j

[0093] Polynomial kernel: K(x) i ,x j )=(αx i T x j +c) d

[0094] Gaussian radial basis core:

[0095] It's understandable that kernel functions enable linear regression or classification in high-dimensional spaces, thereby improving model performance. Kernel ridge regression, support vector regression, and Bayesian nonlinear regression (such as Gaussian process regression) are all effective regression models for solving nonlinear problems. Kernel ridge regression prevents overfitting through regularization, support vector regression improves model robustness by maximizing the margin, while Bayesian nonlinear regression excels in complex problems by providing uncertainty estimation and flexible nonlinear modeling capabilities.

[0096] For example, Kernel Ridge Regression (KRR) is a regression analysis method that combines ridge regression and kernel methods. Its basic idea is to introduce a regularization term to prevent overfitting. Given a training dataset... The loss function for kernel ridge regression is:

[0097]

[0098] Where w is the regression coefficient and λ is the regularization parameter. By differentiating the loss function and setting it to zero, we obtain the closed-form solution for the regression coefficients:

[0099] w=(K+λI) -1 y (4)

[0100] Where K is the kernel matrix calculated by the kernel function, I is the identity matrix, and y is the output vector.

[0101] For example, Support Vector Regression (SVR) is an extension of Support Vector Machine (SVM) designed to solve regression problems. SVR minimizes prediction error by finding a hyperplane with the maximum margin in a high-dimensional feature space. The goal of SVR is to minimize the following loss function:

[0102]

[0103] Where, ξ i and Here, C is the slack variable, and C is the penalty parameter. SVR introduces an ε-insensitive loss function, meaning that the model's error within a certain range will not be penalized.

[0104]

[0105] It is understandable that by introducing a kernel function, SVR can perform linear fitting in high-dimensional space, thereby enhancing the complexity and fitting ability of the model.

[0106] For example, Bayesian regression is a regression method based on probability inference. To handle nonlinear problems, Bayesian regression can be combined with kernel methods or Gaussian process regression. This embodiment is a Bayesian nonlinear regression based on Gaussian processes, which can effectively model complex nonlinear relationships and provide estimates of uncertainty. Gaussian process regression (GPR) is a typical implementation of Bayesian nonlinear regression. The Gaussian process hypothesis function assumes that the output follows a multivariate normal distribution with a mean of zero, and the covariance is defined by the kernel function. By modeling the function, the Gaussian process regression model can handle highly nonlinear regression tasks. The Gaussian process hypothesis is as follows:

[0107]

[0108] Where m(x) is the mean function and k(x,x′) is the covariance function. Commonly used kernel functions include the Radial Basis Function Kernel (RBF).

[0109] Given training data X and output y, and a new input x * The predicted distribution of Gaussian process regression is:

[0110]

[0111] Where, μ * and These are the predicted mean and variance, respectively:

[0112]

[0113]

[0114] For example, the basic nonlinear regression model is a regression analysis defined based on a tree model. The tree model method is a nonparametric approach widely used in classification and regression problems. It constructs a decision tree by recursively partitioning the data space, thereby creating a set of rules for predicting new data. The process of constructing a decision tree model typically includes the following steps:

[0115] i. Data partitioning: At each node, select a feature x i And the corresponding threshold t, divide the dataset into two parts, satisfying x i ≤t and x i >t.

[0116] ii. Selecting the optimal split point: This involves optimizing a metric (e.g., Gini impurity, information gain, or variance reduction) to choose the best features and split point. For regression tasks, minimizing the mean squared error (MSE) is most commonly used to determine the split point.

[0117]

[0118] Among them, y i It is the actual value. This is a predicted value.

[0119] iii. Recursive partitioning: Repeat the above steps on each child node until the stopping condition is met, such as when the number of samples in the node is less than the preset value or the maximum tree depth is reached.

[0120] ⅵ Prediction: For regression tasks, the final predicted value is usually the average of the target values ​​of the samples in the leaf nodes.

[0121] The advantage of tree models lies in their ease of interpretation and handling of complex nonlinear relationships. Ensemble learning methods based on tree models, such as random forests and gradient boosting trees, can reduce overfitting when dealing with high-dimensional data.

[0122] For example, Random Forest Regression (RFR) is an ensemble learning method that improves the generalization ability of a model by constructing multiple decision trees and combining their predictions. Its core idea is "bagging," which involves constructing multiple different decision trees through bootstrap sampling of the training data and averaging the predictions of these trees. The specific steps of Random Forest Regression are as follows:

[0123] i. Data sampling: For each tree, a subset is obtained by randomly sampling with replacement from the training set.

[0124] ii. Feature randomness: During the segmentation process at each node, a subset of features is randomly selected to determine the optimal segmentation point. This increases the diversity of the model and reduces the risk of overfitting.

[0125] iii. Construct decision trees: Construct each tree using the methods described above until a stopping condition is met, such as tree depth or the number of leaf node samples.

[0126] ⅵ Ensemble prediction: The final prediction result is the average of the prediction results from all trees.

[0127]

[0128] Where T is the number of trees in the forest. This is the prediction result for the t-th tree.

[0129] For example, Gradient Boosting Trees (GBT) is an ensemble learning method based on an additive model. It builds multiple decision trees progressively, with each tree optimizing the previous one. Its core idea is to minimize the loss of the objective function to gradually improve the model's prediction accuracy. The specific steps of Gradient Boosting Trees are as follows:

[0130] i. Initial Prediction: Use a simple model (usually a constant value) for initial prediction. Let the initial model be...

[0131] ii. Calculate the residuals: For each sample i, calculate the prediction residuals of the current model:

[0132] r i =y i -f m-1 (x i (13)

[0133] iii. Constructing a decision tree: Training a new decision tree h based on the residuals. m (x) to minimize the loss function Negative gradient direction:

[0134]

[0135] vi. Update Model: Update the model to a weighted sum of the previous model and the new tree:

[0136] f m (x)=f m-1 (x)-ηh(x) (15)

[0137] Where η is the learning efficiency, used to control the contribution of each tree to the final model.

[0138] v. Iteration: Repeat steps ii to vi until the predetermined number of iterations is reached or the error converges.

[0139] Random forest regression and gradient boosting trees are both based on tree model methods and are representative algorithms of ensemble learning. Random forest reduces overfitting by averaging the results of multiple decision trees, while gradient boosting trees improve model accuracy by progressively reducing residuals.

[0140] For example, a basic nonlinear regression model is defined based on regression analysis using a neural network. Each layer in the neural network performs some simple mathematical operations, such as weighted summation and the application of activation functions. These operations can be combined to form complex nonlinear relationships, such as multilayer perceptron regression based on a neural network.

[0141] For example, Multilayer Perceptron Regression (MLP Regression) is a regression model based on neural networks. A Multilayer Perceptron (MLP) is a feedforward neural network that contains at least one hidden layer. MLP regression aims to perform complex pattern learning and regression prediction on input data through layers of nonlinear transformations, exhibiting strong nonlinear modeling capabilities. The structure of MLP regression is as follows:

[0142] i. Input layer: Accepts an n-dimensional input vector x = [x1, x2, ... xn] n Each node represents a feature in the input data.

[0143] ii. Hidden Layer: A multilayer perceptron contains at least one hidden layer. Each neuron in the hidden layer calculates a weighted sum based on the input and weights, and then applies an activation function for a non-linear transformation. The output of each neuron is:

[0144]

[0145]

[0146] Where σ is the activation function, and These are weights and biases, respectively.

[0147] iii. Output Layer: The final output layer is the prediction result. For regression problems, the output layer typically does not have an activation function and directly outputs a continuous value.

[0148]

[0149] in and b (2) These represent the weights and biases of the output layer, respectively, and m is the number of neurons in the hidden layer.

[0150] ⅵ Training process: Calculate the predicted value through forward propagation Then, the prediction error is calculated based on the loss function (such as mean squared error). Next, the gradient of the error with respect to each weight and bias is calculated using the backpropagation algorithm, and the weights and biases are updated using gradient descent.

[0151]

[0152] Where η is the learning rate, L is the loss function, and l represents the network layer.

[0153] S202, Hyperparameter Optimization.

[0154] In this embodiment, six nonlinear regression analysis methods are selected: kernel ridge regression, support vector regression, Bayesian nonlinear regression, random forest regression, gradient boosting tree, and multilayer perceptron regression. These methods are used to define the search range of hyperparameters for each basic nonlinear regression model. A hyperparameter grid is then constructed based on the search range to find the optimal hyperparameter combinations for each basic nonlinear regression model. Finally, the hyperparameter combinations of each basic nonlinear regression model are evaluated to determine the target hyperparameter combination for each model. Table 1 shows the search range and optimal parameters for the defined hyperparameters of the nonlinear regression models.

[0155]

[0156] It should be noted that the methods for finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid include grid search, random search, and Bayesian optimization methods. By combining these methods, we can flexibly choose the appropriate method based on different models and data characteristics, ensuring optimal performance with less computational resources.

[0157] For example, a combination of grid search and cross-validation is used to find the hyperparameter combinations for each basic nonlinear regression model. By combining parameters one by one in a preset hyperparameter grid, the model performance is comprehensively evaluated, which is suitable for cases with a small number of parameters.

[0158] For example, training with randomly selected hyperparameter combinations in a hyperparameter grid based on random search is used to find the hyperparameter combinations for each basic nonlinear regression model. Training with randomly selected hyperparameter combinations is more efficient in time than grid search, and is particularly suitable for high-dimensional hyperparameter spaces.

[0159] For example, a surrogate model is constructed based on Bayesian optimization methods, and a getter function is used for iterative selection to find the hyperparameter combination for each basic nonlinear regression model. Hyperparameters are iteratively selected by constructing a surrogate model (e.g., a Gaussian process) and a getter function (e.g., the desired improvement). The optimization process includes an initialization phase, where several initial points are randomly selected in the hyperparameter space and their performance is evaluated; in the iterative optimization phase, the surrogate model is updated based on existing hyperparameter-performance data, and the getter function is used to select the next optimal hyperparameter point; finally, the optimization terminates when the maximum number of iterations is reached or the model performance no longer significantly improves. In addition, genetic algorithms and particle swarm optimization are also introduced as optimization methods. The former simulates the crossover and mutation of parameters through natural selection, while the latter uses swarm intelligence to simulate the movement of particles in the search space.

[0160] For example, the negative mean squared error is used to evaluate the hyperparameter combinations of each underlying nonlinear regression model being searched, thereby determining the target hyperparameter combination.

[0161] as follows:

[0162]

[0163] Where, n cv For the cross-validation fold number (5 folds), n i Let be the amount of data in the i-th fold.

[0164] S300. Screen the basic nonlinear regression models that have undergone hyperparameter optimization and obtain the basic nonlinear regression models that are not overfitted.

[0165] In this embodiment, the hyperparameter combination of the basic nonlinear regression model is obtained through hyperparameter optimization, so as to determine whether the basic nonlinear regression model is overfitted. The overfitted basic nonlinear regression model is discarded, while the nonfitted basic nonlinear regression model is retained.

[0166] In some embodiments, see Figure 3 , Figure 3 This is a flowchart of the model evaluation and screening method provided in the embodiments of this disclosure, where S300 includes S301 to S305:

[0167] S301. Calculate multiple performance indicators for each basic nonlinear regression model.

[0168] In this embodiment, the various performance indicators include the following:

[0169] 1) Mean Squared Error (MSE):

[0170]

[0171] 2) Mean Absolute Error (MAE):

[0172]

[0173] 3) Coefficient of Determination (R²) 2 ):

[0174]

[0175] 4) Root Mean Squared Error (RMSE):

[0176]

[0177] 5) Mean Absolute Percentage Error (MAPE):

[0178]

[0179] Where n is the number of samples, y i This is the actual value. This is a predicted value.

[0180] For example, as shown in Table 2,

[0181]

[0182]

[0183]

[0184] MSE train Training set mean squared error, The training set determination coefficients are shown in Table 2, and the MSE... cv Mean Squared Error of Cross-Validation (MSE) i Let be the mean squared error of the i-th fold. Table 2 shows the statistics of relevant indicators for model evaluation and screening:

[0185]

[0186] S302. Set the difference threshold between the training set and cross-validation based on each performance metric.

[0187] In this embodiment, for each performance metric, a threshold value is set for the difference between the training set and cross-validation, such as δ. MSE δ MAE And so on. A decision criterion is set: if the model meets the following conditions on all metrics, it is considered not overfitting and is retained:

[0188]

[0189] Where the subscript cv represents cross-validation and the subscript train represents the training set.

[0190] S303. Determine and screen basic nonlinear regression models that are not overfitted based on the difference threshold of the training set and cross-validation.

[0191] In this embodiment, as exemplarily shown in Table 2, the MSE difference is calculated as follows:

[0192] ΔMSE=|MSE cv -MSE train | (29)

[0193] If ΔMSE≤5, the model is considered not overfitted and is retained.

[0194] The calculation results are shown in Table 2. It can be seen that Kernel Ridge Regression, Support Vector Regression, Random Forest Regression, Gradient Boosting Tree, and Multilayer Perceptron Regression meet the screening criteria and are retained for the next step of dynamic weight calculation.

[0195] S304. Conduct a diversity assessment of the basic nonlinear regression model that is not overfitted and obtain the correlation of the model residuals.

[0196] In this embodiment, for any two models M i and M j Calculate the Pearson correlation coefficient of the prediction residuals on the validation set:

[0197]

[0198] in,

[0199] Based on the correlation of the residuals of the calculation models, models with low residual correlation are selected first to increase the effectiveness of model fusion.

[0200] S305. Sort the basic nonlinear regression models that are not overfitted according to the correlation of the model residuals.

[0201] In this embodiment, by ranking the basic nonlinear regression models, the models with better performance can be selected first, which is beneficial for the weighted fusion of models.

[0202] S400, based on a nonlinear regression model that has not overfitted, performs dynamic weighted fusion.

[0203] In this embodiment, dynamic weights are calculated for the selected basic nonlinear regression models based on multiple performance indicators; adaptive fusion is then performed on the selected basic nonlinear regression models based on the results of the dynamic weight calculation. The optimal fusion strategy is selected based on model performance and diversity; weighted average, stacked regression, or voting methods can be used to make the calculated values ​​of the fused models closer to the true values.

[0204] For example, the weighting formula for the above performance indicators is as follows:

[0205]

[0206] Among them: Metric ij Let γ be the value of the i-th model on the j-th performance metric. j Let be the weight coefficient of the j-th performance index, satisfying ρ ik For model Mi With model M k The residual correlation coefficient between them, δ is the weight coefficient of the diversity index, and ε is the smoothing factor to prevent the denominator from being zero.

[0207] Therefore, the calculated weights are normalized, and the normalized weights are ω. i ′:

[0208]

[0209] For example, as shown in Table 3, we select MSE and R here. 2 The formula for calculating the weights is as follows:

[0210]

[0211] Where, ω i Let be the weights of the i-th model;

[0212] α and β are adjustment coefficients that satisfy α + β = 1, where α = 0.5 and β = 0.5.

[0213] ε is a smoothing factor to prevent the denominator from being zero or other numerical problems; here, ε = 0.001.

[0214] MSE i and The training sets MSE and R of the i-th model are respectively. 2 .

[0215] And for ω i The final weight ω is obtained by normalizing the weights using equation (32). i ′, as shown in Table 3 - combining MSE and R 2 The dynamically adjusted model weights are shown below:

[0216]

[0217]

[0218] For example, the specific performance of weighted average, stacked regression, or voting method is as follows:

[0219] Weighted average: When the models have similar performance and high diversity, weighted average fusion is used.

[0220] Stacked regression: Stacked regression is used when there are large differences in model performance and high diversity.

[0221] Voting method: For classification tasks or regression tasks that require discretized output, the voting method can be considered.

[0222] Formula for selecting fusion strategy:

[0223]

[0224] Where, σ perf The standard deviation of the model performance index. For the average residual correlation, σ perf , This is a preset threshold.

[0225] S500 trains the fused basic nonlinear regression model to obtain the target nonlinear regression model, and obtains shale oil reservoir parameters based on the target nonlinear regression model.

[0226] In this embodiment of the disclosure, an adaptive fusion strategy using stacked regression is adopted to be suitable for situations where there are large differences in model performance and high diversity. By constructing a stacked regressor, the basic nonlinear regression model can be trained and learned to form a target nonlinear regression model. The final output value of the target nonlinear regression model is the shale oil reservoir parameter. The shale oil reservoir parameter can be calculated based on the hyperparameters calculated in the above steps, and then input drilling, logging and well logging data to calculate the parameter.

[0227] In some embodiments, see Figure 4 , Figure 4 This is a flowchart of the dynamic weighted fusion method provided in the embodiments of this disclosure, where S500 includes S501 to S504:

[0228] S501. Based on stacked regression, the selected basic nonlinear regression models are weighted to form a base model.

[0229] In this embodiment, the set of basic nonlinear regression models {M1, M2, ..., M} after screening and weighting is... k} as the base model.

[0230] S502. Select one of the various basic nonlinear regression models as the meta-learner based on the base model.

[0231] In this embodiment, an appropriate meta-learner is selected according to the adaptive fusion strategy. This meta-learner can be one of the following: linear regression, ridge regression, Lasso regression, decision tree regression, support vector regression, or other regression models. In this embodiment, ridge regression is selected as the meta-learner, and its loss function is:

[0232]

[0233] Where λ is the regularization parameter, and w = [w1, w2, ..., w k ] represents the weight vector.

[0234] Based on this, a stacked model is trained. The prediction matrix is ​​obtained by predicting from the training set using the base model set.

[0235] F = [M1(X) train M2(X) train )...M k (X train (36)

[0236] Using a meta-learner on the prediction matrix F and the true value y train Training is performed on the [topic name] to optimize the weights w.

[0237] S503. Train the stacked model based on the selected meta-learners to obtain the target nonlinear regression model.

[0238] In this embodiment, a first-level prediction matrix is ​​generated:

[0239] F = [M1(X) train M2(X) train ) … M k (X train (37)

[0240] Then use F and y train Training M meta Training the meta-learner.

[0241] S504. Perform fusion uncertainty estimation on the target nonlinear regression model.

[0242] In this embodiment, the following methods can be used: Gaussian process regression: The prediction variance of the Gaussian process regression model is used to provide an uncertainty estimate. Monte Carlo Dropout: In the deep learning model, the Monte Carlo Dropout method is used to perform multiple forward propagations to calculate the mean and variance of the prediction results. Confidence interval calculation: Based on the prediction variance, the confidence interval of the prediction results is calculated.

[0243] S600: Evaluate and update the model based on the target nonlinear regression model after training.

[0244] In this embodiment, for model evaluation, various performance metrics of the fused target nonlinear regression model can be calculated on the test set based on the base model, such as MSE, MAE, and R. 2 Then, based on the calculation results, the target nonlinear regression model can be compared with a single model among various basic nonlinear regression models, and the superiority of the fused target nonlinear regression model can be verified based on the comparison results. Regarding online learning and model updates, online learning algorithms are designed for real-time or streaming data; when new data arrives, the model parameters and structure are dynamically adjusted to maintain the model's timeliness.

[0245] For example, test set prediction:

[0246] The base model makes predictions on the test set, resulting in a prediction matrix:

[0247] F test =[M1(X test M2(X) test )...M k (X test (38)

[0248] Using the trained meta-learner, for F test Make a prediction and obtain the final predicted value.

[0249] For example, the mean squared error (MSE) and coefficient of determination (R²) of the test set are calculated. 2 ):

[0250]

[0251]

[0252] MSE test Training set mean squared error, Training set determination coefficient This represents the average of the target values ​​for the test set.

[0253] For example, the results are compared: the performance of the stacked regressor is compared with that of each individual model, and it is found that the test set MSE of the stacked regressor after dynamic weighted fusion is the smallest, at 4.84778, verifying the superiority of the method of the present invention (as shown in Table 4). Simultaneously, the calculated porosity data is loaded into Resform software, and it is found that the calculated porosity and the measured porosity have a good agreement, further illustrating that the porosity calculated based on the dynamic weighted nonlinear regression model has good application effects, such as... Figure 5 As shown), Figure 5 A schematic diagram illustrating porosity calculated using a dynamically weighted nonlinear regression model, provided for embodiments of this disclosure. Table 4 - Performance comparison between the stacked regressor and individual models:

[0254]

[0255] Based on the above method, this disclosure also provides a reservoir characterization parameter calculation system corresponding to the above method, such as... Figure 6 As shown, Figure 6The block diagram of the oil reservoir characterization parameter calculation system provided in this embodiment of the disclosure includes a data preprocessing module 210, a model definition and hyperparameter module 220, a model screening module 230, a dynamic weighted fusion module 240, a training module 250, and an evaluation and update module 260.

[0256] The data preprocessing module 210 is used to acquire datasets of engineering and geological parameters and preprocess the datasets. The model definition and hyperparameter module 220 is used to define multiple basic nonlinear regression models based on the preprocessed dataset and optimize the hyperparameters of these models. The model selection module 230 is used to select the optimized basic nonlinear regression models and obtain those that are not overfitted. The dynamic weighted fusion module 240 is used to perform dynamic weighted fusion based on the selected basic nonlinear regression models. The training module 250 is used to train the fused basic nonlinear regression models to obtain a target nonlinear regression model, and then obtain shale oil reservoir parameters based on the target nonlinear regression model. The evaluation and update module 260 is used to evaluate the trained target nonlinear regression model and perform online learning and model updates based on the evaluated target nonlinear regression model.

[0257] The aforementioned system comprehensively considers the nonlinear relationships of data, multiple performance indicators, and model diversity. It performs hyperparameter tuning, dynamically adjusts model weights and fusion strategies, integrates uncertainty estimation, and provides online learning and model update mechanisms. This results in a high-performance, highly generalizable, and interpretable regression model. It can effectively handle large and complex shale oil datasets, improving the accuracy and reliability of predictions and providing a solid scientific basis for shale oil exploration and development. It has high practical value and significant potential for wider application.

[0258] Based on the same inventive concept as the above-disclosed content, this disclosure also provides an electronic device 3000. For example... Figure 7 As shown, Figure 7 This is a block diagram of an electronic device provided in an embodiment of the present disclosure. The electronic device 3000 of this embodiment includes at least one processor 310 and at least one memory 320 electrically connected to it. The memory 320 is electrically connected to the processor 310, and the memory 320 stores instructions that can be executed by at least one processor 310. The instructions are executed by at least one processor 310 to enable at least one processor 310 to perform the method described above.

[0259] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. The indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.

[0260] Based on the same inventive concept, this disclosure also provides a computer storage medium storing a computer program, which, when executed by a processor, is described as described above. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0261] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for calculating shale oil reservoir parameters, characterized in that, The method includes: Obtain a dataset of engineering parameters and geological parameters, and preprocess the dataset; Based on the preprocessed dataset, several basic nonlinear regression models are defined, and hyperparameter optimization is performed on these models. Screen the basic nonlinear regression models that have undergone hyperparameter optimization and obtain the basic nonlinear regression models that are not overfitted; Dynamic weighted fusion is performed based on a nonlinear regression model that has not overfitted. The fused basic nonlinear regression model is trained to obtain the target nonlinear regression model, and shale oil reservoir parameters are obtained based on the target nonlinear regression model.

2. The method according to claim 1, characterized in that, The preprocessing of the dataset containing the engineering and geological parameters includes: The dataset includes a training set of the engineering parameters and the geological parameters, and a test set of the engineering parameters and the geological parameters; The ratio of the training set to the test set is one of 6:4, 7:3, 8:2 or 9:

1.

3. The method according to claim 1, characterized in that, The process of defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to these models includes: The basic nonlinear regression model is defined based on a kernel function, which includes: Given two samples x i and x j The kernel function K is defined as follows: K(x i ,x j )=φ(x i ) T φ(x j ) Where φ is a nonlinear mapping function.

4. The method according to claim 3, characterized in that, The various basic nonlinear regression models defined based on kernel functions include: kernel ridge regression, support vector regression, and Bayesian nonlinear regression.

5. The method according to claim 1, characterized in that, The process of defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to these models includes: The basic nonlinear regression model tree model is defined, and the tree model includes: Divide the data, and at each node, select a feature x. i And the corresponding threshold t, divide the dataset into two parts, satisfying x i ≤t and x i >t; Selecting the optimal split point involves optimizing the metric to choose the best features and split point: Among them, y i It is the actual value. It is a predicted value; Recursively partition the tree, repeating the above steps on each child node until a stopping condition is met, such as when the number of samples in a node is less than a preset value or when the maximum tree depth is reached. The prediction is the average of the target values ​​of the samples in the leaf nodes.

6. The method according to claim 5, characterized in that, The various basic nonlinear regression models described are defined based on tree models, including random forest regression and gradient boosting tree.

7. The method according to claim 1, characterized in that, The process of defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to these models includes: The basic nonlinear regression model is defined based on a neural network; The neural network is defined by a model including multilayer perceptron regression.

8. The method according to claim 7, characterized in that, The multilayer perceptron regression includes: The input layer accepts an n-dimensional input vector x = [x1, x2, ... xn]. n Each node represents a feature in the input data; Hidden layers: A multilayer perceptron contains at least one hidden layer. Each neuron in the hidden layer calculates a weighted sum based on the input and weights, and applies an activation function for a non-linear transformation. The output of each neuron is: Where σ is the activation function, and These are weights and biases, respectively. The output layer ultimately outputs the prediction result, and outputs a continuous value: in and b (2) These are the weights and biases of the output layer, respectively, and m is the number of neurons in the hidden layer. During the training process, predicted values ​​are calculated through forward propagation. Then, the prediction error is calculated based on the loss function; next, the gradient of the error with respect to each weight and bias is calculated using the backpropagation algorithm, and the weights and biases are updated using gradient descent. Where η is the learning rate, L is the loss function, and l represents the network layer.

9. The method according to claim 1, characterized in that, The process of defining multiple basic nonlinear regression models based on the preprocessed dataset and optimizing hyperparameters according to these models includes: To determine the search range of hyperparameters for each basic nonlinear regression model; Construct a hyperparameter grid based on the search range; Find the hyperparameter combination for each basic nonlinear regression model based on the hyperparameter grid; The evaluation is performed based on the hyperparameter combination of each basic nonlinear regression model to determine the target hyperparameter combination of each basic nonlinear regression model.

10. The method according to claim 9, characterized in that, The methods for finding hyperparameter combinations for each basic nonlinear regression model based on a hyperparameter grid include grid search, random search, and Bayesian optimization methods.

11. The method according to claim 10, characterized in that, The process of finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: finding the hyperparameter combination of each basic nonlinear regression model based on the grid search combined with cross-validation.

12. The method according to claim 10, characterized in that, The step of finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: training a set of hyperparameters randomly selected from the hyperparameter grid based on the random search to find the hyperparameter combination of each basic nonlinear regression model.

13. The method according to claim 10, characterized in that, The process of finding the hyperparameter combination of each basic nonlinear regression model based on the hyperparameter grid includes: constructing a surrogate model based on the Bayesian optimization method and obtaining the iterative selection of the function to find the hyperparameter combination of each basic nonlinear regression model.

14. The method according to claim 9, characterized in that, The evaluation based on the hyperparameter combinations of each basic nonlinear regression model determines the target hyperparameter combinations for each basic nonlinear regression model, including: The negative mean square error is used to evaluate the hyperparameter combinations of each basic nonlinear regression model to determine the target hyperparameter combination.

15. The method according to claim 1, characterized in that, The process of screening the hyperparameter-optimized basic nonlinear regression models and obtaining basic nonlinear regression models that are not overfitting includes: Calculate multiple performance metrics for each basic nonlinear regression model; Set a threshold for the difference between the training set and cross-validation based on each performance metric; Based on the training set and the difference threshold of cross-validation, the underlying nonlinear regression models that are not overfitted are determined and screened.

16. The method according to claim 15, characterized in that, The various performance indicators include: mean square error, mean absolute error, coefficient of determination, root mean square error, and mean absolute percentage error.

17. The method according to claim 15, characterized in that, The process of screening the hyperparameter-optimized basic nonlinear regression models and obtaining basic nonlinear regression models that are not overfitting includes: Diversity assessment of the underlying nonlinear regression model that is not overfitted, and correlation of the model residuals are obtained; The underlying nonlinear regression models that are not overfitted are ranked based on the correlation of the model residuals.

18. The method according to claim 15, characterized in that, The dynamic weighted fusion based on the nonlinear regression model that has not overfitted includes: Dynamic weight calculation is performed on the selected basic nonlinear regression model based on the aforementioned multiple performance indicators; The selected basic nonlinear regression models are adaptively fused based on the dynamic weight calculation results.

19. The method according to claim 18, characterized in that, The strategy for adaptively fusing the selected basic nonlinear regression models based on the dynamic weight calculation results includes weighted average, stacked regression, and voting method.

20. The method according to claim 2, characterized in that, The step of training and learning the fused multiple basic nonlinear regression models to obtain a target nonlinear regression model, and obtaining shale oil reservoir parameters based on the target nonlinear regression model, includes: Based on the stacked regression, the selected basic nonlinear regression models are weighted to form a base model; Based on the base model, one of several basic nonlinear regression models is selected as the meta-learner; The target nonlinear regression model is obtained by stacking the selected meta-learners and training the model. Shale oil reservoir parameters are obtained based on the target nonlinear regression model.

21. The method according to claim 20, characterized in that, The method further includes: The evaluation is based on the target nonlinear regression model after training and learning; Online learning and model updates are performed based on the evaluated target nonlinear regression model.

22. The method according to claim 21, characterized in that, The evaluation based on the trained target nonlinear regression model includes: Based on the base model, calculate various performance indicators of the fused target nonlinear regression model on the test set; Based on the calculation results, the target nonlinear regression model is compared with a single model among a variety of basic nonlinear regression models; The superiority of the fused target nonlinear regression model is verified based on the comparison results.

23. A shale oil reservoir parameter calculation system, characterized in that, The system includes: The data preprocessing module is used to acquire datasets of engineering parameters and geological parameters, and to preprocess the datasets. The model definition and hyperparameter module is used to define multiple basic nonlinear regression models based on the preprocessed dataset, and to optimize the hyperparameters of the multiple basic nonlinear regression models. The model screening module is used to screen basic nonlinear regression models that have undergone hyperparameter optimization and to obtain basic nonlinear regression models that are not overfitted. The dynamic weighted fusion module is used to perform dynamic weighted fusion based on the selected basic nonlinear regression models. The training module is used to train and learn the fused basic nonlinear regression model to obtain the target nonlinear regression model, and to obtain shale oil reservoir parameters based on the target nonlinear regression model.

24. An electronic device, characterized in that, The electronic device includes: Memory, used to store computer programs; A processor, configured to execute the computer program to implement a method for calculating reservoir characterization parameters as described in any one of claims 1 to 22.

25. A computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements a method for calculating oil reservoir characterization parameters as described in any one of claims 1 to 22.

Images