A method for predicting formation physical property parameters based on a machine learning optimization model

Abstract

The present application relates to the technical field of geological parameter analysis, and particularly relates to a stratum physical property parameter prediction method based on a machine learning optimization model, comprising the following steps: obtaining well logging curves of multiple wells, and calculating stratum porosity and permeability of the multiple wells; obtaining distribution characteristics of the well logging curves, the porosity and the permeability, and analyzing correlations of the well logging curves, the porosity and the permeability; dividing the multiple wells into blind wells and non-blind wells, and selecting data in well logging curve, porosity and permeability data of the non-blind wells as a test set and a training set; constructing and optimizing a machine learning optimization model for stratum physical property parameter prediction on the test set and the training set, and evaluating applicability of the machine learning optimization model through the blind wells. The present application utilizes powerful nonlinear mapping capability of machine learning, establishes porosity and permeability directly predicted according to well logging curves, avoids complicated formula operation process, and realizes a substantial improvement in prediction automation degree.

Description

Technical Field

[0001] This invention relates to the field of geological parameter analysis technology, specifically to a method for predicting formation physical parameters based on a machine learning optimization model. Background Technology

[0002] In the exploration and development of oil and gas and natural gas hydrate resources, accurately obtaining formation physical parameters, especially porosity and permeability, is a core step in reservoir evaluation, reserve calculation, and exploitation scheme optimization. Traditional methods for obtaining formation physical parameters mainly rely on two types of methods: one is through core experiments, which, although highly accurate, is costly, time-consuming, and limited in core sampling, making it difficult to obtain continuous profiles; the other is based on calculations using classic well logging interpretation models (such as the Wyllie-Rose and Timur formulas). These models are usually based on simplified geological assumptions and have poor applicability and low prediction accuracy in reservoirs with complex lithology and strong heterogeneity (such as natural gas hydrate reservoirs).

[0003] Therefore, developing an intelligent prediction method that can automatically predict, has high accuracy, and is highly applicable is of great significance for improving the accuracy and efficiency of reservoir physical parameter prediction. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting formation physical parameters based on a machine learning optimization model, so as to solve the technical problems of low prediction accuracy, poor applicability and low degree of automation in the existing technology.

[0005] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0006] A method for predicting formation physical parameters based on a machine learning optimization model includes the following steps:

[0007] Obtain logging curves from multiple wells, and calculate formation porosity and permeability from multiple wells;

[0008] Obtain the distribution characteristics of logging curves, porosity, and permeability, and analyze the correlation between logging curves, porosity, and permeability;

[0009] Multiple wells were divided into blind wells and non-blind wells, and data from the logging curves, porosity and permeability data of non-blind wells were selected as test set and training set;

[0010] A machine learning optimization model for predicting formation physical parameters was built and optimized on the test and training sets, and the applicability of the machine learning optimization model was evaluated through blind wells.

[0011] As a preferred embodiment of the present invention, the logging curve includes natural gamma, resistivity, sonic transit time, density, and well diameter.

[0012] As a preferred embodiment of the present invention, the method for calculating porosity includes:

[0013] Calculation of clay content in sandstone and conglomerate using natural gamma-ray logging values Where GR is the natural gamma logging value, These are natural gamma-ray logging values ​​from pure sandstone sections. The natural gamma logging value is for the pure mudstone section;

[0014] right Nonlinear correction yields , where C is an empirical coefficient related to stratigraphic age;

[0015] Will Porosity obtained by transformation ,in This is the density logging value. The skeletal density of pure sandstone. The density of pure mudstone, ρ represents the pore fluid density.

[0016] As a preferred embodiment of the present invention, the method for calculating the permeability includes:

[0017] Through porosity Corrected mud content Establishing permeability through wettability experience relationship Where a, b, and c are empirical coefficients related to lithology;

[0018] Substituting the empirical values ​​of a, b, and c, we get... , among which when At that time, take ,when ,Pick .

[0019] As a preferred embodiment of the present invention, the method for obtaining the distribution characteristics includes:

[0020] The distribution frequencies of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were statistically analyzed separately.

[0021] Plot a distribution map with distribution frequency as the vertical axis and natural gamma, resistivity, acoustic transit time, density, well diameter, porosity, and permeability as the horizontal axis;

[0022] Based on the distribution map, the main distribution ranges and peak characteristics of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were determined.

[0023] As a preferred embodiment of the present invention, the correlation analysis method includes:

[0024] The correlation coefficients between pairs of data in natural gamma, resistivity, sonic transit time, density, well diameter, porosity and permeability are calculated using the Pearson correlation coefficient, forming a correlation coefficient matrix.

[0025] Plot scatter plots between pairs of data for natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability to form a joint distribution map.

[0026] As a preferred embodiment of the present invention, the method for constructing the machine learning optimization model includes:

[0027] 5000 sets of data were selected from the logging curves, porosity and permeability data of non-blind wells, and 80% of them were used as the training set and 20% as the test set.

[0028] Three machine learning models for predicting porosity and permeability were obtained by training three models—Support Vector Machine (SVM), Random Forest (RF), and Distributed Neural Network (DNN)—based on the training set.

[0029] The Sparrow Search Algorithm (SSA) is used to evaluate the performance of three machine learning models based on the error of the validation set and iterates to optimize the hyperparameters of the three machine learning models, resulting in three optimized machine learning models.

[0030] As a preferred embodiment of the present invention, the hyperparameters of the Sparrow Search Algorithm (SSA) for optimizing the Support Vector Machine (SVM) include a penalty coefficient. Kernel function type, kernel function coefficients ;

[0031] The hyperparameters of the Sparrow Search Algorithm (SSA) that optimize Random Forest (RF) include: the number of decision trees. Limit the number of features considered in the branch, the minimum number of nodes required for the intermediate node branch, and the maximum depth F of the decision tree;

[0032] The hyperparameters of the Sparrow Search Algorithm (SSA) optimized for a Deep Neural Network (DNN) include the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate. Batch size and number of iterations.

[0033] As a preferred embodiment of the present invention, the method for evaluating the applicability of the machine learning optimization model includes:

[0034] Based on the logging curves of blind wells, the goodness of fit R between the predicted and actual values ​​of porosity and permeability of three machine learning optimization models was calculated. 2 ;

[0035] The goodness of fit R of porosity and permeability was calculated using three machine learning optimization models. 2 Conduct an applicability assessment.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] This invention utilizes the powerful nonlinear mapping capabilities of machine learning to establish a method for directly predicting porosity and permeability based on well logging curves, avoiding cumbersome formula calculations and significantly improving the automation of prediction.

[0038] This invention employs the Sparrow Search algorithm to automatically optimize the hyperparameters of various machine learning models in parallel, avoiding the blindness and inefficiency of traditional manual trial-and-error parameter tuning. This ensures that the model always runs under optimal or near-optimal configuration, resulting in a significant improvement in prediction accuracy. Attached Figure Description

[0039] To more clearly illustrate the embodiments of the present invention or the technical solutions in 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 merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0040] Figure 1 A flowchart of a method for predicting formation physical parameters based on a machine learning optimization model provided in an embodiment of the present invention;

[0041] Figure 2 Well logging curves and porosity-permeability distribution characteristics in the Muli area provided in this embodiment of the invention;

[0042] Figure 3 The well logging curves, porosity-permeability distribution combined distribution maps, and correlations provided in the embodiments of the present invention;

[0043] Figure 4 A comparison diagram of formation porosity and permeability predicted by the SVM model and the actual formation porosity and permeability provided in this embodiment of the invention;

[0044] Figure 5 A comparison diagram of formation porosity and permeability predicted by the RF model and the actual formation porosity and permeability provided in the embodiments of the present invention;

[0045] Figure 6 A comparison diagram of formation porosity and permeability predicted by the DNN model and the actual formation porosity and permeability provided in this embodiment of the invention;

[0046] Figure 7 This is a ranking diagram of the contribution of feature parameters provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, this invention provides a method for predicting formation physical parameters based on a machine learning optimization model, comprising the following steps:

[0049] Obtain logging curves from multiple wells, and calculate formation porosity and permeability from multiple wells;

[0050] Obtain the distribution characteristics of logging curves, porosity, and permeability, and analyze the correlation between logging curves, porosity, and permeability;

[0051] Multiple wells were divided into blind wells and non-blind wells, and data from the logging curves, porosity and permeability data of non-blind wells were selected as test set and training set;

[0052] A machine learning optimization model for predicting formation physical parameters was built and optimized on the test and training sets, and the applicability of the machine learning optimization model was evaluated through blind wells.

[0053] The logging curves include natural gamma, resistivity, sonic transit time, density, and well diameter.

[0054] This invention uses the calculation results (porosity, permeability) of classical physical models as training targets for machine learning, ensuring that the prediction results have clear physical meaning. Simultaneously, it leverages the powerful nonlinear mapping capabilities of machine learning to overcome the insufficient accuracy of traditional empirical formulas in complex reservoirs, achieving a significant improvement in accuracy, as detailed below:

[0055] The method for calculating the porosity includes:

[0056] Calculation of clay content in sandstone and conglomerate using natural gamma-ray logging values Where GR is the natural gamma logging value, These are natural gamma-ray logging values ​​from pure sandstone sections. The natural gamma logging value is for the pure mudstone section;

[0057] right Nonlinear correction yields , where C is an empirical coefficient related to stratigraphic age;

[0058] Will Porosity obtained by transformation ,in This is the density logging value. The skeletal density of pure sandstone. The density of pure mudstone, ρ represents the pore fluid density.

[0059] The method for calculating the permeability includes:

[0060] Through porosity Corrected mud content Establishing permeability through wettability experience relationship Where a, b, and c are empirical coefficients related to lithology;

[0061] Substituting the empirical values ​​of a, b, and c, we get... , among which when At that time, take ,when ,Pick .

[0062] This invention does not directly use machine learning to regress physical property parameters. Instead, it uses the calculation results of classical rock physics models (such as the clay-corrected density-porosity model and the Timur permeability model) as training labels for machine learning. This allows the "black box" capabilities of machine learning to be built on solid physical laws, ensuring the geological rationality of the prediction results while using machine learning to compensate for the shortcomings of traditional models under complex conditions.

[0063] The methods for obtaining the distribution characteristics include:

[0064] The distribution frequencies of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were statistically analyzed separately.

[0065] Plot a distribution map with distribution frequency as the vertical axis and natural gamma, resistivity, acoustic transit time, density, well diameter, porosity, and permeability as the horizontal axis;

[0066] Based on the distribution map, the main distribution ranges and peak characteristics of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were determined.

[0067] The correlation analysis method includes:

[0068] The correlation coefficients between pairs of data in natural gamma, resistivity, sonic transit time, density, well diameter, porosity and permeability are calculated using the Pearson correlation coefficient, forming a correlation coefficient matrix.

[0069] Plot scatter plots between pairs of data for natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability to form a joint distribution map.

[0070] Before modeling, this invention mandates data distribution characteristic analysis and Pearson correlation analysis, generating distribution maps and joint distribution maps. This analysis identifies outliers and clarifies variable relationships (such as discovering a strong negative correlation between density and porosity), providing crucial insights for subsequent feature engineering and model interpretation, thus avoiding blind modeling.

[0071] The method for constructing the machine learning optimization model includes:

[0072] 5000 sets of data were selected from the logging curves, porosity and permeability data of non-blind wells, and 80% of them were used as the training set and 20% as the test set.

[0073] Three machine learning models for predicting porosity and permeability were obtained by training three models—Support Vector Machine (SVM), Random Forest (RF), and Distributed Neural Network (DNN)—based on the training set.

[0074] Among them, the kernel functions of Support Vector Machines (SVM) mainly fall into three categories: polynomial kernel functions, radial basis functions (RBF), and sigmoid functions. This invention uses the radial basis function (RBF), whose expression is: .

[0075] The core approach of using SVM to solve problems encountered in regression fitting has shifted from finding the optimal plane to finding an optimal surface that minimizes the error between all original data and the target classification surface.

[0076] Generally, we assume the dataset contains l training samples. ,in , , This is the corresponding output value.

[0077] Let the regression polynomial established in the new space be: ;

[0078] in, Let ε be a linear loss function in a multinomial nonlinear function mapping expression.

[0079] ;

[0080] Where f(x) is the predicted value of the objective function; y is the true value. This means that if the difference between the objective expression f(x) and the true value y is less than or equal to ε, then the value of the loss function is equal to 0.

[0081] Slack variables were introduced The mathematical expression for SVM is:

[0082] ;

[0083] Where C is the penalty value, the larger C is, the larger the penalty function is for samples with training error greater than ε, and ε represents the upper limit of the objective function error. The smaller ε is, the smaller the objective function error is.

[0084] Use the Largerange function to solve for the dual form of the above equation:

[0085] ;

[0086] in, This is the kernel function.

[0087] Let the optimal solution obtained by the above equation be Then we can obtain the expressions for w* and b*. Therefore, the regression function is: .

[0088] Random Forest (RF), proposed by Leo Breiman in 2001, is an improvement on Bagging and consists of a large number of decision trees. RF can meet the needs of classification and regression. For classification, voting determines the prediction result of each decision tree, and the class with the most votes is selected. For regression, the mean of all decision trees is selected as the prediction result.

[0089] The RF regression model can be mathematically interpreted as follows: Given a specific dataset, X is the independent variable (input data), and Y is the dependent variable to be predicted (output data). Assuming that the distributions of (X,Y) are independent, a training set is randomly generated from (X,Y), and the prediction result is denoted as g(X). Then, its mean squared generalization error is expressed as: ;

[0090] Assuming there are h decision trees, what is the predicted value of h decision trees? The average value is the prediction result of random forest regression. If Then the following formula holds: ;

[0091] In the above formula The generalization error is represented by PE. ** When h is infinitely large, the average generalization error of a single decision tree is denoted as PE. * PE * satisfy: ;

[0092] The above equation satisfies: ;

[0093] in This represents the weighted correlation coefficient between the residuals. The final regression function for the randomization function is: ;

[0094] The construction steps or process of this algorithm are as follows:

[0095] Step 1: Use the Bootstrap resampling method to randomly select n samples with replacement from the dataset, repeating this sampling K times to obtain K training sample sets, which are independent and identically distributed. Simultaneously, the samples not selected in the K samplings are grouped into K out-of-bag (OOB) data sets, becoming the K test sample sets.

[0096] Step 2: Construct a decision tree using the K training samples extracted in Step 1. A random subspace method is used to select the node feature variables of the decision tree; that is, when a sample has n features, m features are randomly selected from each node of the decision tree. Then, the information values ​​contained in these features are calculated, and the most representative feature and its corresponding value are selected from the m features for node splitting.

[0097] Step 3: Each tree will grow to its maximum extent until the number of samples in a node is less than the originally set threshold or the minimum mean square error is less than the set threshold, at which point the decision tree terminates.

[0098] Step 4: Based on the above steps, K decision trees were obtained. Combining these K decision trees can construct a random forest regression model. Then, inputting the data will yield... ,Will average As a prediction result of the random forest regression model.

[0099] After generating pre-pruned decision trees, feature importance can be obtained using the Gini index of each decision tree in the random forest. Let G be the Gini index, and if dataset A has M classes in a classification problem, the probability that a subset belongs to the Mth class is... Then the Gini index is: ;

[0100] In a decision tree, the change in the Gini index before and after pruning a certain internal node is denoted as V. Suppose that after pruning, two new nodes appear, and the Gini indexes of these two new nodes are denoted as . Then we have the following formula: ;

[0101] Let feature N be in the decision tree. Let S be the set of nodes that appear in the decision tree. Then, the feature N in the decision tree... The importance of is recorded as follows: ;

[0102] If there are K decision trees in a random forest, then the Gini score of feature N in the random forest algorithm is: ;

[0103] Finally, after normalizing the Gini indices obtained from different features, the importance ranking of the different features is obtained by sorting them.

[0104] Neural networks are products derived from the analysis, abstraction, and imitation of the operation of biological neural networks. They are network structures that connect a large number of individual neurons according to certain rules, forming a network capable of processing information in parallel. They belong to a type of intelligent mathematical operation model. This model was first proposed by an American psychologist who analyzed the characteristics of biological neurons and, after combining them with mathematical theory, proposed the first mathematical model of an artificial neuron.

[0105] A neural network consists of a large number of processing units called neurons (corresponding to nerve cells in the human brain). The function of a neuron is to calculate the inner product of the input vector and the weight vector, and then pass it through a nonlinear transfer function to obtain a scalar result.

[0106] ;

[0107] ;

[0108] Commonly used activation functions include:

[0109] (1) Sigmoid function: ;

[0110] (2) tanh function: ;

[0111] (3) ReLU function: ;

[0112] (4) Leaky ReLU function: ;

[0113] Taking a neural network model containing an input layer, two hidden layers, and an output layer as an example, the first... The linear coefficient from the i-th neuron in layer 1 to the j-th neuron in layer 1 is defined as... The bias of the j-th neuron in the l-th layer is defined as The bias of the j-th neuron in the l-th layer is defined as The output of the first hidden layer is calculated as shown in the following equation:

[0114] ;

[0115] Suppose that the activation function we choose is The output of the j-th neuron in the (l+1)-th layer is defined as Then the input to the second hidden layer is:

[0116] ;

[0117] The output calculation for the second hidden layer is shown in the following formula:

[0118] ;

[0119] The input to the output layer is:

[0120] ;

[0121] The output of the output layer is then:

[0122] ;

[0123] The general expression for the loss function of a neural network is: The loss function is used to measure the degree of inconsistency between the true value y and the predicted value f(x), and generally, a smaller value is better. A small loss function indicates that the machine learning model closely approximates the true distribution of the data, indicating good model performance; a large loss function indicates that the machine learning model deviates significantly from the true distribution of the data, indicating poor model performance. For regression problems, the loss function is:

[0124] ;

[0125] in, y represents the predicted result, and y represents the actual result.

[0126] Gradient descent is one of the most commonly used optimization algorithms and currently the most frequently used method for optimizing neural networks. It is an algorithm that reaches the minimum value of a function through iterative steps. To find a local minimum of a function using gradient descent, steps can be taken that are proportional to the negative value of the function's gradient (or approximate gradient) at the current point.

[0127] The gradient represents the directional derivative of a function at a given point, where the directional derivative along that direction reaches its maximum value; in other words, it is the derivative of the function at the current position. ;

[0128] In the formula, As the independent variable, For about The function, Let represent the gradient. Then the independent variable parameter in the i-th step should be equal to: ;

[0129] η is the learning rate. When the learning rate is too low, many steps are needed to converge. Conversely, when the learning rate is too high, gradient descent will fail to reach the minimum value.

[0130] The learning process of BP neural network reservoir modeling consists of two stages: The first stage is forward propagation, which involves inputting known reservoir geological information as training samples. Using the established network structure and the thresholds and weights obtained from the previous training step, the output of each neuron is calculated from the first layer onwards. This process processes the input geological variable data and outputs the results. The second stage is backpropagation, where the output error is propagated back through the hidden layers to the input layers, distributing the error to all units in each layer. This yields the error signal for each unit, which serves as the basis for adjusting the weights. Specifically, by comparing the deviation between the actual output and the expected parameters, the network error is determined and propagated back to the reservoir attribute data processing units, thus readjusting their connections. This process of forward and backward propagation of signals and adjustment of layer weights is repeated continuously. This continuous adjustment of weights is the network's learning and training process. This continues until the network's output error is reduced to an acceptable level or until a predetermined number of training iterations are completed.

[0131] Assumption It is known that the following can be calculated using the chain rule:

[0132] ;

[0133] and ;

[0134] Therefore, as long as the error of the l-th layer is known... Then you can calculate , , , and Used to update the parameters w and b of the l-th layer. It is passed to the (l-1)th layer to further calculate the derivatives of the parameters of the (l-1)th layer.

[0135] The Sparrow Search Algorithm (SSA) is used to evaluate the performance of three machine learning models based on the error of the validation set and iterates to optimize the hyperparameters of the three machine learning models, resulting in three optimized machine learning models.

[0136] The hyperparameters of the Sparrow Search Algorithm (SSA) for optimizing Support Vector Machines (SVM) include the penalty coefficient. Kernel function type, kernel function coefficients ;

[0137] The hyperparameters of the Sparrow Search Algorithm (SSA) that optimize Random Forest (RF) include: the number of decision trees. Limit the number of features considered in the branch, the minimum number of nodes required for the intermediate node branch, and the maximum depth F of the decision tree;

[0138] The hyperparameters of the Sparrow Search Algorithm (SSA) optimized for a Deep Neural Network (DNN) include the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate. Batch size and number of iterations.

[0139] In the Sparrow Search Algorithm (SSA), sparrows can be categorized into three types: leaders, followers, and alerters. Within the population, leaders play a role in finding resources, providing clues such as foraging areas and directions. Followers, with their keen observation skills, capture the leader's movements to obtain food. Sparrows primarily utilize both leader and follower strategies for foraging, maintaining a cooperative yet competitive relationship. Alerters typically reside on the periphery of the population, where they are more vulnerable to predators, thus requiring them to constantly adjust their position to secure safer areas. Based on this brief overview, the following principles can be summarized:

[0140] (1) Resource reserve principle: Leaders usually control more food resources and are responsible for finding more favorable areas for the population, that is, areas with abundant food, and contacting followers to provide relevant foraging areas and directions.

[0141] (2) The principle of constant proportion: As long as a better resource is found, every sparrow can become a leader. However, the overall proportion of the two remains constant. That is to say, for every sparrow whose status changes to a leader, there must be another sparrow whose status changes to a follower. In other words, the status of leaders and followers is in a dynamic switching process.

[0142] (3) Warning and escape principle: If a sparrow discovers a predator, it will issue a warning. Once the warning value exceeds the warning threshold, the followers will escape to other safe areas to forage under the guidance of the leader.

[0143] (4) The principle of surveillance and plunder: When sparrows are foraging, they will always find the discoverer with the best food resources and obtain food from that area or forage around the leader. At the same time, if the followers want to improve their foraging rate, they will monitor the leader in real time in order to plunder food resources.

[0144] (5) The principle of flying towards the best: The scarcer the resources available to followers, the worse their foraging locations will be within the population. Some followers who are starving are more likely to fly to areas with more resources to forage.

[0145] (6) Early warning and transfer principle: When sensing a dangerous atmosphere, the sparrow individuals on the periphery will immediately transfer to a safe position, while the individuals in the middle of the population will adopt a random movement strategy to move closer to other individuals.

[0146] The process of the sparrow search algorithm includes:

[0147] The first step: Establish a population X composed of n sparrows. The expression of population X is:

[0148] ;

[0149] where d is the number of hyperparameters to be optimized, is the position of the nth sparrow in the dth dimension;

[0150] The second step: Set the fitness value represents the ability of the sparrow individual to find food. The fitness value The expression is:

[0151] ;

[0152] where, is the fitness function, is the fitness value of the nth sparrow, is the position vector of the nth sparrow;

[0153] The third step: According to the population X and the fitness value , set the position update formula of the discoverer as:

[0154] ;

[0155] where, is the number of iterations, is the maximum number of iterations, , are respectively at the th, th iteration, the th sparrow as the discoverer at the rd dimension position, is a random number in [0,1]. is the early warning value, and the value range is [0,1], is the safety value, and the value range is [0.5,1]. Q is a random number obeying the normal distribution, and L is a 1×d matrix, and each element is 1;

[0156] When the situation of R2 < ST occurs, it means that the foraging environment is safe at this time and no predators are found, and the discoverer can perform a large number of search operations;

[0157] When R2>ST occurs, it means that an individual in the sparrow population has already discovered the predator and sounded the alarm. At this time, all sparrows will take anti-predation behavior and quickly move to other safe places to continue foraging.

[0158] Step 4: Based on population X and fitness value The location update formula for discoverers is set as follows:

[0159] ;

[0160] in, In the first In the next iteration, the sparrow, as the discoverer, is in the best position. Let A be the worst possible position for the sparrow that discovers the sparrow, and let A be a 1×d matrix where each element is randomly assigned a value of 1 or -1. ;

[0161] When it appears In this case, it indicates that the entrant with low fitness has not found food in the current state and needs to go to other areas to forage in order to improve its survival rate.

[0162] Step 5: Set the number of scout sparrows to 10% to 20% of the total population. Their initial positions are randomly generated within the population. The behavioral expression for scout alerts is:

[0163] ;

[0164] in, Let t be the current globally optimal sparrow position in the t-th iteration, β be a control parameter, a random number following a standard normal distribution, K be a uniform random number with a value range of [-1, 1], and ε be a small constant to avoid the denominator being zero; The fitness value of the sparrow acting as a scout. The sparrow in the best position in the whole game fitness value, Sparrow in the worst position globally The fitness value.

[0165] When f appears i =f g In this situation, it indicates that the sparrow in question is in the optimal position globally and will move closer to other sparrows to reduce the risk of being preyed upon.

[0166] When f appears i >f g In this situation, it indicates that the sparrow in question is on the edge of the population and is more vulnerable to predators, and will quickly move to the current optimal position.

[0167] The methods for optimizing the hyperparameters of the Sparrow Search algorithm are as follows:

[0168] Represent the position of each sparrow as a (penalty coefficient). Kernel function type, kernel function coefficients ) combination or (number of decision trees) The constraints are the number of features considered in the branch, the minimum number of nodes required for intermediate branches, and the maximum depth of the decision tree (F) combined with (the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate). The initial position is randomly generated within a preset range, using a combination of batch size and iteration steps, and is expressed as a relative standard error (RSEM) or correlation coefficient (R). 2 The fitness function is used to calculate the fitness value of the population. The population size and maximum number of iterations are set for the Sparrow Search Algorithm (SSA).

[0169] During the iteration process of the Sparrow Search Algorithm (SSA):

[0170] Update the position of the leading sparrow according to the discoverer's update formula, and explore new (penalty coefficient) Kernel function type, kernel function coefficients ) combination or (number of decision trees) The constraints are the number of features considered in the branch, the minimum number of nodes required for intermediate branches, and the maximum depth of the decision tree (F) combined with (the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate). Combinations of batch size and number of iterations;

[0171] The position of the following sparrow is updated according to the formula for updating the position of the introducer, so as to move closer to the high-quality solution;

[0172] The position of the early warning sparrow is updated based on the behavior formula of the reconnaissance and early warning system to avoid getting trapped in local optima.

[0173] The fitness values ​​of all sparrows are recalculated after each iteration;

[0174] The position of the sparrow with the highest fitness that reaches the maximum number of iterations is used as the penalty coefficient. Kernel function type, kernel function coefficients () or (number of decision trees) The constraints include the number of features considered in the branch, the minimum number of nodes required for intermediate branches, and the maximum depth of the decision tree (F) or (the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate). The optimal combination solution of batch size and number of iterations.

[0175] This invention proposes a framework for simultaneously optimizing the hyperparameters of three models—SVM, RF, and DNN—using SSA, and objectively comparing them using the R² metric on blind wells. This ensures that each model reaches its maximum potential. Multi-model comparison avoids the limitations of a single model, and a competition mechanism ensures that the final solution is optimal for the current task. R² provides a clear and quantifiable evaluation criterion.

[0176] This invention explicitly classifies well locations into training wells (non-blind wells) and validation wells (blind wells), and uses the model's predictive performance (R²) on blind wells as the core criterion for final evaluation and model selection. It simulates the scenario of "inferring the unknown from the known" in real exploration, effectively preventing the model from overfitting to training wells and ensuring that the selected model possesses strong extrapolation capabilities and practical application value.

[0177] The methods for evaluating the applicability of the machine learning optimization model include:

[0178] Based on the logging curves of blind wells, the goodness of fit R between the predicted and actual values ​​of porosity and permeability of three machine learning optimization models was calculated. 2 ;

[0179] The goodness of fit R of porosity and permeability was calculated using three machine learning optimization models. 2 Conduct an applicability assessment.

[0180] Example:

[0181] The Muli Depression is located in the upper reaches of the Datong River basin in the central Qilian Mountains. Its southern part lies within the thrust fault zone on the northern edge of the Datong Mountains, while its northern part is situated on the fault zone on the southern edge of the Tuolai Mountains. The overall structural morphology is a narrow, elongated compound syncline. Statistical analysis was conducted on logging parameters such as natural gamma, apparent resistivity, sonic transit time, density, and well diameter of the hydrate reservoirs in the Muli permafrost area, and logging curves for wells DK2, DK3, and SK2 were plotted.

[0182] Porosity is one of the important parameters for reservoir evaluation. One of the basic tasks in well logging evaluation of formations is to accurately determine the porosity value of the reservoir. At present, the main methods for determining the porosity of natural gas hydrate reservoirs using well logging include density logging (DEN), neutron logging (CNL), acoustic logging (AC), resistivity logging (Rt), and nuclear magnetic resonance logging (NMR). In well sections with high hydrate content, Lin Zhenzhou et al. [4] suggested using density values ​​to calculate porosity. Natural gas hydrates are usually hosted in argillaceous sandstone, so the presence of argillaceous material will inevitably affect the well logging response value. In order to estimate the magnitude of its influence, the argillaceous content in sandstone and conglomerate is usually calculated by the following equation using natural gamma logging values. :

[0183] ;

[0184] Where GR is the natural gamma logging value. These are natural gamma-ray logging values ​​from pure sandstone sections. This represents the natural gamma ray logging value for a pure mudstone section. The above formula also requires nonlinear correction; the correction formula is:

[0185] ;

[0186] Where C is an empirical coefficient related to the age of the strata, and 3.7 is taken for new strata (Neogene strata). The hydrate reservoirs encountered by drilling in the Muli area are mainly Neogene strata.

[0187] When transformed into an expression for the porosity of natural gas hydrate reservoirs, it can be simplified to:

[0188] ;

[0189] In the actual calculation process, This is the density logging value. , The mudstone section was taken as having a value of 2.3 g / cm³. .

[0190] (2) Permeability is used to quantitatively describe the permeability of reservoir rocks and has important guiding significance for the exploration, development, and production of oil and gas fields. Due to the heterogeneity of reservoirs and the complexity of geological conditions, predicting permeability is extremely difficult. The Timur formula is used to calculate permeability, that is: ;

[0191] in To bind water saturation, % is porosity, %; K is absolute permeability, 10⁻³ μm.

[0192] Bound water in rocks includes stagnant water in microcapillary pores, stagnant water at the bends of small channels in other capillary pores, and thin film stagnant water on the surface of hydrophilic rock particles. Therefore, the degree of bound water saturation necessarily depends on the geometry of the rock pores and the magnitude of capillary forces, the clay content of the rock, and its wettability. In predominantly hydrophilic sandstone, the rock specific surface area is a comprehensive reflection of these factors; the larger the specific surface area, the higher the bound water saturation. Reservoir systems containing hydrates have complex structures, requiring necessary assumptions to simplify the reservoir structure and estimate hydrate saturation. The uncertainty in the estimation results also stems from these assumptions. Therefore, to improve interpretation accuracy, multiple methods or reservoir evaluation models should be selected for calculation as much as possible, depending on the actual situation. An empirical relationship for Swb is established using porosity, clay content, and wettability.

[0193] ;

[0194] Where a, b, and c are empirical coefficients related to lithology. A typical expression is:

[0195] ;

[0196] Among them when At that time, take ,if ,Pick .

[0197] The porosity and permeability of the formations in the Muli area were calculated using formulas, and their distribution characteristics are similar to those of well logging data. Figure 2 As shown in the figure, the well diameter exhibits multiple peaks, mainly distributed around the 100mm and 220mm peaks, with a relatively complex distribution pattern; the density logging also shows multiple peaks, with the main peak at approximately 2.3g / cm³. 3 Some samples have a density value less than 2 g / cm³. 3 The data are as follows: Acoustic transit time is mainly distributed between 200 and 600 μm / s, with two peaks, the main peak located near 350 μm / s; natural gamma curves are mainly distributed between 0 and 150 API, with three peaks, the main peak located around 90 API; resistivity is distributed between 0 and 50000 Ω·m, mainly between 0 and 10000 Ω·m, with a large range; porosity is mainly below 30%, with a small amount of data distributed between 30% and 90%; permeability is distributed between 0 and 40000 mD, mainly between 0 and 2000 mD, with a large range.

[0198] The correlation between well logging curves and porosity and permeability is as follows: Figure 3 As shown. Porosity has a weak positive correlation with resistivity and acoustic transit time, but the correlation does not exceed 0.5, and a very strong negative correlation with density, reaching -0.95. Permeability has a weak positive correlation with resistivity, reaching 0.3, and a strong negative correlation with density, reaching -0.69. Porosity has a strong positive correlation with permeability, reaching 0.74. Density has a weak negative correlation with acoustic transit time and resistivity.

[0199] Density and porosity exhibit a good linear correlation, while permeability shows two distinct power-law relationships with both density and porosity. The distributions of porosity, permeability, density, sonic transit time, natural gamma, well diameter, and resistivity are complex, exhibiting multiple peaks.

[0200] Besides having a strong correlation with density, porosity and permeability have weak correlations with other logging curves, making it difficult to establish a good mathematical model.

[0201] Five thousand sets of data were selected from the logging and porosity / permeability data of wells DK3 and SK2 as training and validation sets. The selected data were representative, focusing on the middle parts of each lithological interval. Well DK2 was used as a blind well to test its applicability.

[0202] First, three models—SVM, RF, and DNN—were trained using the training set data. The SSA algorithm evaluated the performance of the machine learning models based on the error on the validation set and iterated to optimize the hyperparameters of the SVM, RF, and DNN models. 80% of the samples were randomly selected from 5000 sets of well logging data for training the models in the SSA algorithm, with the remaining 20% ​​used as the validation set. The SSA algorithm iterated 100 times, with a population size of 80. The range and optimal results of the SSA algorithm's optimized model hyperparameters are shown in Table 1.

[0203] Table 1. Range and Results of Hyperparameters for SSA Algorithm Optimization of Machine Learning Algorithm

[0204]

[0205] During the optimization of hyperparameters using the SSA algorithm, the change in the fitness value reveals that the RF model has the smallest fitness value, the DNN model has a larger fitness value, and the SVM model has the largest. The fitness value decreases with increasing iteration count. The SSA algorithm exhibits a simpler change in fitness value and a longer stabilization time when optimizing the SVM model, while showing more significant changes when optimizing the RF and DNN models.

[0206] Optimized SVM, RF, and DNN models were used to predict the porosity and permeability of well DK2. The predicted formation porosity and permeability were compared with the actual formation porosity and permeability. Figure 4 (SVM model) Figure 5 (RF model) and Figure 6 As shown in the (DNN model), the optimized SVM model has the best prediction performance, with the predicted porosity and permeability showing the best R-squared value compared to the actual reservoir porosity and permeability. 2 All values ​​exceeded 0.9, with an average of 0.9297 (as shown in Table 2). The R-values ​​of the predicted porosity and the actual reservoir porosity from the optimized RF model were... 2 Ideally, it reaches 0.9756, while the Rw of the predicted porosity versus the actual reservoir porosity is... 2 The worst result was only 0.8308, with R0 representing porosity and permeability. 2 The average value is 0.9032. The R-squared values ​​of the predicted porosity and permeability by the optimized DNN model compared to the actual reservoir porosity and permeability are... 2 All exceeded 0.9, with an average of 0.9120. From R...2 On average, the SVM model is the best, followed by the DNN model, and the RF model is the worst. In terms of training time, the SVM algorithm is the shortest, the RF model is the longest, and the DNN model is the longest. Considering both time cost and accuracy, the SVM model is the best, followed by the DNN model, and the RF model is the worst.

[0207] Table 2 Comparison of Evaluation Parameters for SSA Optimization Algorithm

[0208]

[0209] SVM, RF, and DNN models predict porosity better than permeability. Porosity is relatively evenly distributed along the diagonal and is relatively concentrated. The intersection plot of SVM model's predicted permeability and the actual permeability is distributed on both sides of the diagonal but is scattered. Compared with the actual permeability, the permeability predicted by SVM is too high when it is less than 7500mD, and too low when it is not. The permeability predicted by RF model and DNN model is mostly distributed in the upper part of the diagonal, indicating that the permeability predicted by the two models is too high.

[0210] The feature parameters obtained from the random forest algorithm are ranked by their contribution as follows: Figure 7 As shown, the most significant influence of formation porosity and permeability on well logging curves is density, exceeding 0.9, followed by the natural gamma ray curve, while the effects of sonic transit time, well diameter, and depth are minimal. The primary reason for the significant influence of the density logging curve is that porosity and permeability are calculated based on density logging. The natural gamma ray logging response is mainly caused by the formation's natural radioactivity, such as potassium isotopes (K40), which are most common in clay minerals, and the infilling of clay minerals has a significant impact on formation porosity and permeability.

[0211] This invention utilizes the powerful nonlinear mapping capabilities of machine learning to establish a method for directly predicting porosity and permeability based on well logging curves, avoiding cumbersome formula calculations and significantly improving the automation of prediction.

[0212] This invention employs the Sparrow Search algorithm to automatically optimize the hyperparameters of various machine learning models in parallel, avoiding the blindness and inefficiency of traditional manual trial-and-error parameter tuning. This ensures that the model always runs under optimal or near-optimal configuration, resulting in a significant improvement in prediction accuracy.

[0213] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for predicting formation physical parameters based on a machine learning optimization model, characterized in that, Includes the following steps: Obtain logging curves from multiple wells, and calculate formation porosity and permeability from multiple wells; Obtain the distribution characteristics of logging curves, porosity, and permeability, and analyze the correlation between logging curves, porosity, and permeability; Multiple wells were divided into blind wells and non-blind wells, and data from the logging curves, porosity and permeability data of non-blind wells were selected as test set and training set; A machine learning optimization model for predicting formation physical parameters was built and optimized on the test set and training set, and the applicability of the machine learning optimization model was evaluated through blind wells. Methods for calculating permeability include: Through porosity Corrected mud content Establishing permeability through wettability experience relationship Where a, b, and c are empirical coefficients related to lithology; Substituting the empirical values ​​of a, b, and c, we get... , among which when At that time, take ,when ,Pick ; Methods for obtaining distribution characteristics include: The distribution frequencies of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were statistically analyzed separately. Plot a distribution map with distribution frequency as the vertical axis and natural gamma, resistivity, acoustic transit time, density, well diameter, porosity, and permeability as the horizontal axis; Based on the distribution map, the main distribution ranges and peak characteristics of natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability were determined.

2. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 1, characterized in that: The logging curves include natural gamma, resistivity, sonic transit time, density, and well diameter.

3. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 2, characterized in that: The method for calculating the porosity includes: Calculation of clay content in sandstone and conglomerate using natural gamma-ray logging values Where GR is the natural gamma logging value. These are natural gamma-ray logging values ​​from pure sandstone sections. The natural gamma logging value is for the pure mudstone section; right Nonlinear correction yields , where C is an empirical coefficient related to stratigraphic age; Will Porosity obtained by transformation ,in This is the density logging value. The skeletal density of pure sandstone. The density of pure mudstone, ρ represents the pore fluid density.

4. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 3, characterized in that: The correlation analysis method includes: The correlation coefficients between pairs of data in natural gamma, resistivity, sonic transit time, density, well diameter, porosity and permeability are calculated using the Pearson correlation coefficient, forming a correlation coefficient matrix. Plot scatter plots between pairs of data for natural gamma, resistivity, sonic transit time, density, well diameter, porosity, and permeability to form a joint distribution map.

5. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 4, characterized in that: The method for constructing the machine learning optimization model includes: 5000 sets of data were selected from the logging curves, porosity and permeability data of non-blind wells, and 80% of them were used as the training set and 20% as the test set. Three machine learning models for predicting porosity and permeability were obtained by training three models—Support Vector Machine (SVM), Random Forest (RF), and Distributed Neural Network (DNN)—based on the training set. The Sparrow Search Algorithm (SSA) is used to evaluate the performance of three machine learning models based on the error of the validation set and iterates to optimize the hyperparameters of the three machine learning models, resulting in three optimized machine learning models.

6. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 5, characterized in that: The hyperparameters of the Sparrow Search Algorithm (SSA) for optimizing Support Vector Machines (SVM) include the penalty coefficient. Kernel function type, kernel function coefficients ; The hyperparameters of the Sparrow Search Algorithm (SSA) that optimize Random Forest (RF) include: the number of decision trees. Limit the number of features considered in the branch, the minimum number of nodes required for the intermediate node branch, and the maximum depth F of the decision tree; The hyperparameters of the Sparrow Search Algorithm (SSA) optimized for a Deep Neural Network (DNN) include the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the number of nodes in the third hidden layer, and the learning rate. Batch size and number of iterations.

7. The method for predicting formation physical parameters based on a machine learning optimization model according to claim 5, characterized in that: The methods for evaluating the applicability of the machine learning optimization model include: Based on the logging curves of blind wells, the goodness of fit R between the predicted and actual values ​​of porosity and permeability of three machine learning optimization models was calculated. 2 ; The goodness of fit R of porosity and permeability was calculated using three machine learning optimization models. 2 Conduct an applicability assessment.

Images