SSA-LSSVM reservoir lithology fine prediction method based on lithology granularity factor construction
By constructing an SSA-LSSVM model and combining feature selection and discrimination criteria, the problems of overfitting and vertical averaging effects in lithology identification were solved, achieving high-precision prediction of lithological grain size and improving exploration efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN POLYTECHNIC UNIV
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-08
AI Technical Summary
Existing machine learning algorithms are prone to overfitting in lithology identification, making it difficult to achieve precise classification under conditions of scarce core samples. Furthermore, the vertical averaging effect of well logging data leads to distorted lithology judgments, making it difficult to distinguish between thin and thick lithologies.
Texture and gradient features were extracted using sliding window standard deviation statistics and central difference method. Feature selection was performed by combining maximum information coefficient and variance inflation factor. The feature combination was optimized by binary sparrow search algorithm. An SSA-LSSVM model was constructed and a generalized Fisher discrimination criterion was introduced for lithological grain size identification.
It achieves high-precision, automated prediction of lithological grain size under small sample conditions, improves the reliability of oil and gas reservoir evaluation and exploration efficiency, overcomes the information ambiguity caused by instrument averaging effect, and establishes a quantitative grain size discrimination standard.
Smart Images
Figure CN121997741A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of well drilling exploration and analysis, and in particular to a method for precise prediction of reservoir lithology. Background Technology
[0002] Lithological identification is a fundamental and crucial technical task in oil and gas exploration, development, and geological research. In particular, the precise classification of reservoir rock grain sizes (such as coarse sand, medium sand, fine sand, and silt) directly affects the scientific validity of sedimentary facies analysis, reservoir property evaluation (porosity and permeability characteristics), and subsequent oil and gas reservoir description. Currently, the main methods for obtaining stratigraphic lithology and grain size information include core sampling analysis and well logging data interpretation.
[0003] Core sampling analysis, which involves obtaining core samples from underground rock formations through drilling, allows for direct observation and analysis of the lithology and grain size. While it provides the most intuitive and accurate lithology and grain size information, it is costly, time-consuming, and in many practical projects, especially in the early exploration stages, it is difficult to obtain a sufficient number of core samples, thus limiting continuous and comprehensive lithology and grain size evaluation throughout the well section.
[0004] To compensate for the limitations of coring, well logging technology has been widely applied. Well logging curves (such as natural gamma, resistivity, sonic transit time, neutron density, etc.) can continuously reflect the properties of subsurface formations. Existing well logging lithology identification methods mainly include plate intersection, empirical discrimination, and some traditional machine learning algorithms. With the development of big data technology, those skilled in the art have gradually begun to use data mining and machine learning methods for lithology identification, which has reduced the subjectivity of traditional methods to some extent. However, even with the application of machine learning algorithms, existing well logging lithology identification technologies still have significant limitations when facing the need for fine-grained predictions:
[0005] First, the model is prone to overfitting and has poor generalization ability to "grain size" labels. The prediction accuracy of machine learning algorithms heavily depends on the quantity and quality of training samples. In study areas where core samples are scarce, samples with clear "grain size classification" labels (such as distinguishing between fine sand and silt) are extremely rare. This makes machine learning algorithms highly susceptible to overfitting, causing the model to struggle to capture subtle differences in logging responses between different grain size lithologies when predicting uncorked well sections.
[0006] Secondly, the feature processing suffers from a fundamental flaw: it struggles to overcome the grain size confusion caused by the "averaging effect." Existing technologies for processing well logging data have a fundamental cognitive blind spot: they all rely on the "absolute value" of the logging curve (e.g., GR=80 API) as model input. However, when dealing with thin interbedded or gradually changing grain sizes, the vertical resolution limitations of logging instruments cause a "vertical averaging effect" on rapidly changing lithologies, averaging different lithologies. This effect leads to a high degree of overlap (e.g., average GR value) in the logging response values of "thin layers of fine sandstone" and "thick layers of siltstone or mudstone," making them indistinguishable. Therefore, all traditional models that rely on "absolute values" as input inevitably misclassify fine sandstone as siltstone or mudstone, resulting in distorted judgments of the grain size level of high-quality reservoirs.
[0007] In summary, existing machine learning methods rely solely on simple linear analysis in the feature selection stage, failing to effectively capture nonlinear features and redundant information in well logging data. This leads to the model being prone to overfitting under conditions of scarce coring samples and making it difficult to finely classify rock grain size. Summary of the Invention
[0008] To address the aforementioned technical problems, this invention proposes a refined reservoir lithology prediction method based on SSA-LSSVM constructed using a lithological grain size factor. First, sliding window standard deviation statistics and central difference methods are used to extract texture and gradient features from well logging curves, respectively, constructing a high-dimensional engineering feature pool to make explicit the implicit 'grain size heterogeneity' and 'lithological abrupt boundary' of the formation. Second, maximum information coefficient (MIC) and variance expansion factor (VIF) are used for initial feature screening, combined with a binary sparrow search algorithm (BSSA) for wrapper-style refinement, establishing a hybrid feature optimization strategy to effectively eliminate redundant noise and lock in the feature subset most sensitive to grain size. Based on this, a generalized Fisher discriminant criterion is introduced as a fitness function, and the SSA algorithm is used to optimize and construct a 'lithological grain size factor' with the maximum inter-class discriminative power, introducing explicit physical constraints to the model. Finally, an SSA-LSSVM lithological grain size identification model is established, and the SSA algorithm is used to adaptively optimize the penalty factor and kernel function width globally, achieving high-precision and refined prediction of reservoir lithology and grain size.
[0009] To achieve the above objectives, the technical solution of the present invention is implemented as follows:
[0010] A refined reservoir lithology prediction method based on SSA-LSSVM constructed using lithology grain size factor includes the following steps:
[0011] S101: Obtain multi-attribute logging curves for known lithology and grain size classification within the study area, and perform preprocessing;
[0012] S102: Calculate texture features and gradient features using preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool; the texture features are used to characterize lithological grain size heterogeneity, and the gradient features are used to characterize abrupt changes in lithological grain size.
[0013] S103: A hybrid strategy of filtering and wrapping based on binary sparrow search algorithm is adopted to screen the best combination of attribute features that are sensitive to lithological grain size identification from the high-dimensional engineering feature pool;
[0014] S104: Based on the optimal combination of attribute features, construct a lithological grain size factor model. With the goal of maximizing the distinguishability between lithologies of different grain size levels, use the sparrow search algorithm to optimize the weights of the lithological grain size factor model, obtain the optimal lithological grain size factor model, and calculate the lithological grain size factor corresponding to the optimal combination of attribute features.
[0015] S105: Construct an SSA-LSSVM lithology-grain size identification model. The optimal attribute feature combination is fused with the corresponding lithology-grain size factor as input features. Known lithology and grain size classification are used as labels. The sparrow search algorithm is used to optimize the hyperparameters of the least squares support vector machine. The LSSVM lithology-grain size identification model is trained based on the optimal hyperparameters to obtain the trained SSA-LSSVM lithology-grain size identification model.
[0016] S106: Predict the entire study area based on the trained SSA-LSSVM lithological grain size identification model.
[0017] Furthermore, the properties of the logging curves include at least natural gamma, deep lateral resistivity, sonic transit time, compensated neutrons, density, and well diameter;
[0018] Preprocessing includes:
[0019] Invalid values and outlier zero values in the original data were removed, and well logging curves from different sources were depth-aligned and resampled at equal intervals.
[0020] The curve is smoothed using a median filtering algorithm;
[0021] A cubic spline interpolation algorithm is used for fitting and filling to restore the continuous variation trend of the logging response;
[0022] The Z-score standardization method is used to eliminate the dimensional differences and order-of-magnitude effects among various logging attributes, and the data is transformed into a standard normal distribution.
[0023] Furthermore, texture and gradient features are calculated using the preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool, including:
[0024] Based on the sliding window technique, the standard deviation of the depth points of the original logging curve within the window is used as the center depth point of the window. The window is moved point by point along the depth to form a texture feature curve for characterizing the heterogeneity of lithological grain size.
[0025] The central difference method is used to calculate the first derivative of the original logging curve at depth as a gradient feature, forming a gradient feature curve for characterizing abrupt changes in lithological grain size.
[0026] Construct a high-dimensional engineering feature pool that includes the original well logging curves, texture feature curves, and gradient feature curves.
[0027] Furthermore, a hybrid filtering-wrapping strategy based on the binary sparrow search algorithm is employed to screen optimal feature combinations sensitive to lithological grain size identification from the high-dimensional engineering feature pool, including:
[0028] Preliminary selection by filtering method: Based on the maximum information coefficient, a preliminary selection set of features is obtained from the high-dimensional engineering feature pool. Based on the variance inflation factor test, a final selection set of features is obtained from the preliminary selection set by filtering method.
[0029] Wrap-up selection: The features selected by the filtering method are encoded into binary feature vectors. The classification accuracy of the LSSVM model is used as the first fitness function. The binary sparrow search algorithm is used to select the best combination of attribute features from the feature set selected by the filtering method.
[0030] Furthermore, based on the maximum information coefficient, a preliminary set of selected features is obtained from the high-dimensional engineering feature pool, including:
[0031] The maximum information coefficient algorithm is used to calculate the nonlinear correlation strength between each feature and rock grain size category in the high-dimensional engineering feature pool. Based on the nonlinear correlation strength, the nonlinear correlation strength is selected. Features with MIC (Micro-Matching Detection) thresholds are selected and removed, while nonlinear correlation strengths are retained. The features of the MIC screening threshold are used as preliminary screening features to obtain the preliminary screening feature set;
[0032] The final filter feature set is obtained from the initial screening features based on the variance inflation factor test, including:
[0033] For the initial screening feature set, a multiple linear regression is performed with the k-th initial screening feature as the dependent variable and the remaining initial screening features as independent variables, and the coefficient of determination of the multiple linear regression is calculated. Based on the coefficient of determination Calculate the variance inflation factor of the k-th preliminary screening feature: Iteratively calculate the variance inflation factor for each initially screened feature, and remove those with variance inflation factors. The VIF screening threshold is used to select features until the variance inflation factor of all initially screened features is less than the VIF screening threshold, thus obtaining the filter-selected feature set.
[0034] Furthermore, the binary sparrow search algorithm is used to select the optimal combination of attribute features from the feature set selected by the filtering method. This includes: constructing a fitness function based on the classification accuracy of LSSVM: dividing the feature set selected by the filtering method into a training set and a validation set, training the LSSVM model using a K-fold cross-validation strategy, using the average classification accuracy of the LSSVM model on the validation set as the first fitness function, using the binary feature vectors of the filtered features as optimization variables, and using the binary sparrow search algorithm for iterative search to obtain the binary feature vectors representing the selected features. The filtered feature combination represented by the binary feature vectors representing the selected features after the iteration is taken as the optimal feature attribute combination.
[0035] Furthermore, the lithological grain size factor model constructed based on the optimal feature combination is as follows:
[0036]
[0037] in, For depth point Lithological grain size factor For the first The best attribute features in depth The value after normalization; For the first The weights of each feature, This represents the optimal number of attribute features.
[0038] Furthermore, a sparrow search algorithm is used to optimize the weights of the lithological grain size factor model to obtain the optimal lithological grain size factor model. This includes: using the weight vector and bias term of the lithological grain size factor model as optimization variables, taking maximizing the inter-class discrimination of coarse sand, medium sand, fine sand, silt, mudstone, and pebble sandstone as the optimization objective, introducing the generalized Fisher discrimination criterion as the second fitness function, and using the sparrow search algorithm to optimize the weight vector and bias term of the lithological grain size factor model to obtain the optimal lithological grain size factor model.
[0039] Furthermore, a sparrow search algorithm is used to optimize the weight vector and bias term of the lithological grain size factor model, including:
[0040] Step 4.1: Using the optimal combination of attribute features as input, calculate the within-class scatter matrix. Inter-class scatter matrix According to the inner scatter matrix and inter-class scatter matrix Define a second fitness function based on the Fisher criterion: , Let be the weight vector to be optimized. It is the minimum value. Indicates transpose;
[0041] Step 4.2: Map the weight vector and bias of the lithological grain size factor model to be optimized to the spatial location of individual sparrows, set the sparrow search algorithm parameters, and initialize the population;
[0042] Step 4.3: Calculate the fitness value of each individual in the population using the second fitness function, select the discoverer and the newcomer according to the fitness value ranking, and record the global optimal position;
[0043] Step 4.3: Update the positions of the discoverer and the joiner according to the position update formula;
[0044] Step 4.4: Randomly select vigilant individuals from the discoverers and participants, and update their positions using the vigilant position update formula;
[0045] Step 4.5: Repeat steps 4.3-4. When the maximum number of iterations is reached or the fitness value no longer increases for several consecutive generations, output the global optimal position.
[0046] Furthermore, predictions are made for the entire study area based on the trained SSA-LSSVM lithological grain size identification model. This includes: preprocessing the target logging curves to be predicted within the study area, calculating texture and gradient features using the S102 method to obtain texture and gradient feature curves, calculating the lithological grain size factor for each depth point using the optimal lithological grain size factor model of S104, generating continuous lithological grain size factor curves, fusing the lithological grain size factor curves with the target logging curves, texture and gradient feature curves as input, and using the trained SSA-LSSVM lithological grain size identification model to identify the specific lithological grain size level corresponding to each depth point.
[0047] The beneficial effects of this invention are as follows:
[0048] First, to address the information ambiguity caused by the instrument's "vertical averaging effect," a "texture and gradient feature" model is constructed using S102 to "make explicit" hidden heterogeneous information. A hybrid feature optimization strategy combining BSSA, filtering, and wrapping methods effectively eliminates redundant noise and extracts the feature subset most sensitive to granularity. Second, to address the poor geological interpretability of traditional models, a novel "lithological granularity factor" with clear physical orientation is constructed. The SSA algorithm is used to optimize the model by maximizing inter-class discriminative power, forcing the model to learn the physical boundaries of different lithological sizes and establishing a quantitative granularity discrimination standard. Finally, the SSA algorithm is used to adaptively optimize the hyperparameters of the LSSVM model globally. Combined with the high-discriminative feature input constructed in the previous steps, this effectively overcomes the defects of easily getting trapped in local optima and overfitting under small sample conditions. This invention ultimately achieves high-precision, automated prediction of fine lithological categories such as coarse sand, medium sand, and fine sand, significantly improving the reliability and exploration efficiency of oil and gas reservoir evaluation. Attached Figure Description
[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 This is a flowchart of lithology prediction according to an embodiment of the present invention.
[0051] Figure 2 This is a heat map showing the correlation between well logging attributes and specific coefficients in an embodiment of the present invention.
[0052] Figure 3 The lithology prediction results of the model in this embodiment of the invention are shown. Detailed Implementation
[0053] 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.
[0054] A fine-grained reservoir lithology prediction method based on SSA-LSSVM constructed using lithological grain size factors, such as... Figure 1 As shown, the steps include:
[0055] S101: Obtain multi-attribute logging curves of known lithology and grain size classification in the study area and perform preprocessing.
[0056] In this embodiment of the application, obtaining multi-attribute logging curves of known lithology and grain size classification within the study area includes:
[0057] Multiple logging curves of known lithology and grain size classification within the study area were obtained. The properties of the logging curves include, but are not limited to, natural gamma (GR), deep lateral resistivity (LLD), sonic transit time (AC), compensated neutron (CNL), density (DEN), and well diameter (CAL).
[0058] In this embodiment of the application, preprocessing is performed, including:
[0059] Data cleaning and alignment: Remove invalid values and outlier zero values from the original data, and perform depth alignment and equal-interval resampling on logging curves from different sources to ensure strict synchronization of multi-attribute data in the depth domain;
[0060] Random noise suppression: The median filtering algorithm is used to smooth the curve, effectively removing instrument random noise while preserving the edge abrupt features of the formation interface to the greatest extent.
[0061] Missing data completion: For local missing data, cubic spline interpolation algorithm is used to fit and fill in the missing data, restoring the continuous trend of well logging response;
[0062] Normalization: The Z-score standardization method is used to eliminate the dimensional differences and order-of-magnitude effects among various logging attributes, and the data is transformed into a standard normal distribution, providing a unified data basis for the subsequent construction of lithological grain size factors.
[0063] S102: Calculate texture features and gradient features using preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool; the texture features are used to characterize lithological grain size heterogeneity, and the gradient features are used to characterize lithological grain size abrupt changes.
[0064] In this embodiment of the application, texture features and gradient features are calculated using preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool, including:
[0065] First, based on the sliding window technique, the standard deviation of the depth points of the original logging curve within the window is used as the center depth point of the window. The window is moved point by point along the depth to form a texture feature curve for characterizing the heterogeneity of lithological grain size.
[0066] Specifically, this invention employs a sliding window technique independently for each original logging curve, setting a depth window containing a predetermined number of sampling points. As the sliding window moves point by point along the depth, the standard deviation of the original logging curve data points within the window is calculated in real time, and this statistical value is assigned to the window center point. All window center points are arranged by depth to form a "texture feature" curve corresponding to each original logging curve, used to characterize the heterogeneity of lithological grain size. That is, a depth sliding window containing N sampling points is set, and for the window center point... Standard deviation of "texture features" at location The calculation formula is as follows:
[0067] ;
[0068] Where, x j μ is the original well logging curve characteristic value of the j-th depth sampling point within the window. i It is the arithmetic mean of the characteristic values of the logging curves of N sampling points within this window.
[0069] Taking the natural gamma (GR) curve as an example, the algorithm sets a geologically significant depth sliding window, for example, 1 meter in size (corresponding to 8 sampling points). This window slides point by point along the entire depth axis of the GR curve (0.125 meters each time). At each depth point, the algorithm calculates the statistical standard deviation of the raw GR values covered by the current window. This calculated "standard deviation" value is assigned to the center depth point of the window, thus forming a completely new "GR texture curve". The value of this "GR texture curve" quantitatively characterizes the changes in grain size microstructure and the strength of sedimentary rhythm within a specific scale (such as 1 meter). This process is repeated for all other S101 raw curves, such as AC, DEN, and LLD, to generate a complete set of "texture feature" curves.
[0070] Furthermore, the central difference method is used to calculate the first derivative of the original logging curve at depth as a gradient feature, forming a gradient feature curve used to characterize abrupt changes in lithological grain size.
[0071] Specifically, the central difference method is used to calculate the gradient feature g(x) at depth point i in each original logging curve. i ), and the gradient features g(x) of all depth points i Arranged by depth, each original logging curve forms a gradient characteristic curve corresponding to a change in lithology and grain size. The formula for calculating the gradient characteristic is:
[0072] ;
[0073] Where, x i+1 and x i-1They are the original log curve values of the points at the depths adjacent to the previous and the next depths of depth i respectively, and d is the depth sampling interval.
[0074] Taking the GR curve as an example again, the algorithm will calculate the first derivative of the GR curve in terms of depth. To ensure the stability of the calculation, preferably, the central difference method is adopted: at any depth point, its gradient value is determined by subtracting the GR value of its previous point from the GR value of its next point and then dividing by twice the depth sampling interval. This process also generates a brand-new "GR gradient curve", and the "peaks" of this curve precisely correspond to the positions where the grain size changes abruptly or the lithologic interfaces in the formation (such as the contact surface between coarse sand and silt). This process will also be repeatedly executed for all other original curves to generate a complete set of "gradient feature" curves.
[0075] Finally, a high-dimensional engineering feature pool including the original log curves, texture feature curves, and gradient feature curves is constructed. It contains all the original curves of S101, all the newly generated "texture feature" curves, and all the newly generated "gradient feature" curves, providing a complete input for subsequent screening of the feature combinations most sensitive to "lithologies of different grain sizes".
[0076] S103: Adopt a filtering-packing hybrid strategy based on the binary sparrow search algorithm (BSSA) to screen the best attribute feature combinations sensitive to lithology grain size identification from the high-dimensional engineering feature pool.
[0077] In the embodiment of the present application, a filtering-packing hybrid strategy based on the binary sparrow search algorithm (BSSA) is adopted to screen the preferred feature combinations sensitive to lithology grain size identification from the high-dimensional engineering feature pool, including:
[0078] The first stage is the "filtering method" primary selection:
[0079] First, use the maximum information coefficient (MIC) algorithm to calculate the non-linear correlation intensity between each feature in the engineering feature pool and the rock grain size category. Based on the non-linear correlation intensity, screen out the features with non-linear correlation intensity < MIC screening threshold and eliminate them, and retain the features with non-linear correlation intensity ≥ MIC screening threshold as the preliminary screening features to obtain a preliminary screening feature set.
[0080] Furthermore, for the preliminary screening feature set, take the k-th preliminary screening feature as the dependent variable and the remaining preliminary screening features as independent variables for multiple linear regression, and calculate the determination coefficient of the multiple linear regression , based on the determination coefficient calculate the variance inflation factor of the k-th preliminary screening feature: ; iteratively calculate the variance inflation factor of each preliminary screening feature , eliminate the variance inflation factor The VIF screening threshold is used to determine the feature set for filtering, until the variance inflation factor of all initially screened features is less than the VIF screening threshold. The variance inflation factor test quantitatively assesses and eliminates redundant features with multicollinearity.
[0081] The second stage is a selection of "wrapping methods":
[0082] First, each feature in the feature set selected by the filtering method is encoded into a binary feature vector through feature encoding, with selected features being 1 and discarded features being 0.
[0083] Furthermore, a fitness function based on LSSVM classification accuracy is constructed. Specifically, the feature set selected by the filtering method is divided into training and validation sets. A K-fold cross-validation strategy is used to divide the training data into K subsets. After each generation of sparrow population updates, the data is dimensionality-reduced using the current binary feature vectors to obtain a dimensionality-reduced dataset containing only the selected features. An LSSVM model is then trained on this dataset. This LSSVM model differs from the later models and is used solely for feature selection. The average classification accuracy of the model on the validation set is calculated as the fitness value for that individual. ,in Let be the number of correctly classified samples in the k-th fold validation. Let be the total number of samples for the k-th fold validation.
[0084] Using the binary feature vectors selected by the filtering method as optimization variables, an iterative search is performed using the Binary Sparrow Search (BSSA) algorithm. In this embodiment, the specific parameters of the BSSA algorithm are set as follows: population size. , Maximum number of iterations Proportion of discoverers The proportion of vigilant individuals Safety threshold The Sigmoid function is used as the transfer function to map continuous position values to selection probabilities. A binary value is generated based on the selection probability. If the random number is less than the selection probability, the selected position is set to 1; otherwise, the position is set to 0. This yields an updated binary feature vector. The goal is to maximize the fitness function described above, and the optimal attribute feature combination is selected by filtering features represented by the selected binary feature vector. :
[0085] ;
[0086] The Include One preferred feature, For the first Each feature vector contains optimized texture features, gradient features, and the original logging curve.
[0087] S104: Construct a lithological grain size factor model based on the optimal combination of attribute features. With the goal of maximizing the distinguishability between lithologies of different grain size levels, use the Sparrow Search Algorithm (SSA) to optimize the weights of the lithological grain size factor model, obtain the optimal lithological grain size factor model, and calculate the lithological grain size factor corresponding to the optimal combination of features.
[0088] In this embodiment of the application, the lithological grain size factor model constructed based on the optimal feature combination is as follows:
[0089] ;
[0090] in, For the first The best attribute features in depth The value after normalization; For the first The fusion weight of each feature, the magnitude of which is positive or negative, quantitatively characterizes the correlation direction between well logging features and lithological grain size variations. When the value is positively correlated with the granularity (the larger the eigenvalue, the coarser the granularity), it indicates that the eigenvalue is positively correlated with the granularity (the larger the eigenvalue, the coarser the granularity). When the value is zero, it indicates that the eigenvalue is negatively correlated with the granularity (the larger the eigenvalue, the finer the granularity). This is the bias term of the model. The lithological grain size factor is an intermediate physical variable that can quantitatively characterize the variation in lithological grain size.
[0091] In this embodiment of the application, the Sparrow Search Algorithm (SSA) is used to optimize the weights of the lithological grain size factor model, obtain the optimal lithological grain size factor model, and calculate the lithological grain size factor of each feature in the optimal feature combination, including:
[0092] Using the weight vector and bias term of the lithological grain size factor model as optimization variables, and maximizing the inter-class discrimination of coarse sand, medium sand, fine sand, silt, mudstone, and pebble sandstone as the optimization objective, the generalized Fisher discrimination criterion is introduced as the fitness function. The Sparrow Search Algorithm (SSA) is used to optimize the weight vector and bias term of the lithological grain size factor model, thereby obtaining the optimal lithological grain size factor model.
[0093] The Sparrow Search Algorithm (SSA) is used to optimize the weight vector and bias term of the lithological grain size factor model, including:
[0094] Step 4.1, define the fitness function;
[0095] To ensure that the constructed lithological grain size factor can distinguish different lithologies to the greatest extent, a generalized Fisher criterion is introduced as the fitness function. The core idea is to find the optimal projection direction. This ensures that, after projection (i.e., the calculated grain size factor), samples of the same type of lithology are clustered as much as possible (small intra-class dispersion), while the centers of different types of lithology are separated as much as possible (large inter-class dispersion).
[0096] Let the preferred feature matrix be... Where N is the total number of samples and m is the number of features; lithological grain size label Corresponding to coarse sand, medium sand, fine sand, silt, mudstone, and pebble sandstone; total number of categories The sample set of the k-th class is The sample size is .
[0097] Step 4.11, calculate the within-class scatter matrix. First, calculate the characteristic mean vector of the k-th lithology. (set up (column vector)
[0098] ;
[0099] in, For the first The feature vector of each depth point (represented as) (column vector).
[0100] Furthermore, calculate the first... Class covariance matrix :
[0101] ;
[0102] Finally, the weighted summation yields the within-class scatter matrix. :
[0103] ;
[0104] Step 4.12, calculate the inter-class scatter matrix. First, calculate the global mean vector of all samples. :
[0105] ;
[0106] Furthermore, the inter-class scatter matrix is calculated. :
[0107] ;
[0108] Step 4.13, Calculate individual fitness: Construct the following maximization objective function. As the fitness value of the SSA algorithm: ;
[0109] in, Let be the weight vector to be optimized. Minimum value (e.g.) This is to prevent the denominator from being zero. The larger the value of this function, the better the grain size factor calculated by the current weight combination can distinguish lithology.
[0110] Step 4.2, Parameter Setting and Population Construction; The parameters of the lithological grain size factor model to be optimized (i.e., feature fusion weights) are set... and bias terms This is mapped to the spatial location of individual sparrows. The parameter is set as the population size. , Maximum number of iterations Discoverer ratio The proportion of vigilant individuals Safety threshold Then, construct the initial population matrix. , of which The position vector of a sparrow is represented as The dimension of the position vector is... (m is the number of preferred features). A random initialization strategy is adopted, within the range of values... The initial population location is generated internally.
[0111] Step 4.3: Assess the fitness of individuals; use the generalized Fisher discriminant criterion function defined in step 5.13. Calculate the fitness value of each sparrow in the population. Based on fitness value Sort the population in descending order; the population with the best (largest) fitness value is ranked first. One individual is marked as a discoverer (responsible for finding food, i.e., exploring the globally optimal solution); the remaining... Each individual is marked as a participant (responsible for following the discoverer, i.e., local development); the current global optimal position is also recorded. and its fitness and global worst fitness .
[0112] Step 4.4, Location Update for Discoverer and Joiners; the discoverer is responsible for finding food in the vast search space. Its location update formula is:
[0113] ;
[0114] in, This represents the current iteration number; Indicates the first The generation The sparrow in the first The position of the dimension; It is a random number; This is a warning value; This is the safety threshold (0.8 in this example). These are random numbers that follow a normal distribution. for A matrix of all 1s.
[0115] Joiners follow the discoverer in search of food; if their energy is too low (lower ranking), they fly elsewhere to forage. The updated formula is: ;
[0116] in, This is the worst position globally at present; The optimal position occupied by the current discoverer; It is A matrix whose elements are randomly 1 or -1, and .
[0117] Step 4.5, the vigilant is updated; randomly selected from the population. Each individual, acting as a watchdog (whether a discoverer or a participant), has the following position update formula:
[0118] ;
[0119] in, This is the current globally optimal position; The step size control parameter follows a normal distribution with a mean of 0 and a variance of 1. for Random numbers; For the current individual fitness, and These represent the current best and worst fitness values globally, respectively. It is a very small constant.
[0120] Step 4.6, Termination Conditions and Factor Output; Determine if the maximum number of iterations has been reached. Check if the fitness value has not increased for several consecutive generations (convergence); if the termination condition is met, output the global optimum position. Analysis Vector, front The position is the optimal feature fusion weight. The last term is the optimal bias term. Substitute the optimal parameters into The formula can be used to calculate the lithological grain size factor at all depth points in the entire area.
[0121] S105: Construct an SSA-LSSVM lithology-grain size identification model. The optimal feature combination is fused with the corresponding lithology-grain size factor as input features. Known lithology and grain size classification are used as labels. The Sparrow Search Algorithm (SSA) is used to optimize the hyperparameters of the Least Squares Support Vector Machine (LSSVM). The LSSVM lithology-grain size identification model is trained based on the optimal hyperparameters to obtain the trained SSA-LSSVM lithology-grain size identification model.
[0122] In this embodiment, an SSA-LSSVM lithological grain size identification model is constructed. First, the optimal feature combination at each depth point is vector-concatenated and fused with the corresponding lithological grain size factor as input features. Known lithologies and grain size classifications (i.e., coarse sand, medium sand, fine sand, silt, mudstone, and pebble sandstone) are used as labels to construct a sample set, and 80% of the sample set is selected as the training set and 20% as the test set.
[0123] Subsequently, the Sparrow Search Algorithm (SSA) was used to apply the penalty factor to the key hyperparameter of the Least Squares Support Vector Machine (LSSVM). and kernel width Adaptive optimization is performed. In this process, the mean squared error (MSE) of the K-fold cross-validation (using 5 folds) of the LSSVM model on the training set is used as the fitness function of SSA, and the optimization objective is to minimize the fitness value. The parameters of SSA are set as follows: population size is
[30] , maximum number of iterations is
[50] , discoverer ratio is [20%], and warning value is set to [0.8]. The LSSVM model is based on the principle of minimizing structural risk. Its loss function consists of the sum of squares of the regularization term and the error term. The optimization problem is transformed into solving a system of linear equations by using the Lagrange multiplier method.
[0124] Finally, based on the optimal penalty factor and kernel function width obtained by SSA optimization, the model is trained using the training set to obtain the final SSA-LSSVM lithological grain size identification model; the model is validated using the test set, with classification accuracy and confusion matrix as the main evaluation criteria to assess the model's identification performance for different lithological grain sizes.
[0125] S106: Predict the entire study area based on the trained SSA-LSSVM lithological grain size identification model.
[0126] In this embodiment, the target logging curve to be predicted in the study area is preprocessed, and then the texture and gradient features are calculated using the method of S102 to obtain the texture feature curve and gradient feature curve. The optimal lithological grain size factor model of S104 is used to calculate the lithological grain size factor at each depth point, generating a continuous lithological grain size factor curve. The lithological grain size factor curve is fused with the target logging curve, texture feature curve and gradient feature curve as input, and then a prediction vector space isomorphic to the training set is constructed. The lithological grain size factor curve is input into the trained SSA-LSSVM lithological grain size recognition model to accurately identify the specific lithological grain size level corresponding to each depth point, such as coarse sand, medium sand, fine sand, silt, mudstone and pebble sandstone. Finally, a continuous lithological profile with fine grain size features is output, realizing the effective characterization of the heterogeneity of the underground reservoir.
[0127] Figure 2 This is a heatmap illustrating the correlation between well logging attributes in an embodiment of the present invention. The heatmap accurately identifies redundant information in the data, i.e., highly correlated attributes. This allows for the simplification and optimization of the input attribute set when constructing a lithology identification model, effectively reducing data dimensionality, avoiding multicollinearity, and significantly improving model computational efficiency and robustness. Simultaneously, the heatmap identifies key well logging attributes with the highest discriminative power for specific lithology classifications, allowing for the priority selection of these most representative features as core inputs, ensuring higher accuracy and interpretability of the constructed lithology prediction model. This method leverages the correlation insights between well logging attributes to significantly improve the intelligence level of lithology identification and the reliability of geological evaluation.
[0128] Figure 3 The image shows the model prediction results of an embodiment of the present invention. As can be seen from the image, different lithology predictions can be made for each depth point on the well logging curve. This point-by-point prediction capability ensures high-precision and continuous characterization of subsurface lithology features, significantly improving the refinement, automation, and reliability of geological interpretation, and providing valuable assistance for subsequent oil and gas resource exploration.
[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A refined reservoir lithology prediction method based on SSA-LSSVM constructed using lithological grain size factors, characterized in that, The steps include: S101: Obtain multi-attribute logging curves for known lithology and grain size classification within the study area, and perform preprocessing; S102: Calculate texture features and gradient features using preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool; the texture features are used to characterize lithological grain size heterogeneity, and the gradient features are used to characterize abrupt changes in lithological grain size. S103: A hybrid strategy of filtering and wrapping based on binary sparrow search algorithm is adopted to screen the best combination of attribute features that are sensitive to lithological grain size identification from the high-dimensional engineering feature pool; S104: Based on the optimal combination of attribute features, construct a lithological grain size factor model. With the goal of maximizing the distinguishability between lithologies of different grain size levels, use the sparrow search algorithm to optimize the weights of the lithological grain size factor model, obtain the optimal lithological grain size factor model, and calculate the lithological grain size factor corresponding to the optimal combination of attribute features. S105: Construct an SSA-LSSVM lithology-grain size identification model. The optimal attribute feature combination is fused with the corresponding lithology-grain size factor as input features. Known lithology and grain size classification are used as labels. The sparrow search algorithm is used to optimize the hyperparameters of the least squares support vector machine. The LSSVM lithology-grain size identification model is trained based on the optimal hyperparameters to obtain the trained SSA-LSSVM lithology-grain size identification model. S106: Predict the entire study area based on the trained SSA-LSSVM lithological grain size identification model.
2. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 1, characterized in that, The properties of well logging curves include at least natural gamma, deep lateral resistivity, sonic transit time, compensated neutrons, density, and well diameter; Preprocessing includes: Invalid values and outlier zero values in the original data were removed, and well logging curves from different sources were depth-aligned and resampled at equal intervals. The curve is smoothed using a median filtering algorithm; A cubic spline interpolation algorithm is used for fitting and filling to restore the continuous variation trend of the logging response; The Z-score standardization method is used to eliminate the dimensional differences and order-of-magnitude effects among various logging attributes, and the data is transformed into a standard normal distribution.
3. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 2, characterized in that, Texture and gradient features are calculated using preprocessed multi-attribute logging curves to construct a high-dimensional engineering feature pool, including: Based on the sliding window technique, the standard deviation of the depth points of the original logging curve within the window is used as the center depth point of the window. The window is moved point by point along the depth to form a texture feature curve for characterizing the heterogeneity of lithological grain size. The central difference method is used to calculate the first derivative of the original logging curve at depth as a gradient feature, forming a gradient feature curve for characterizing abrupt changes in lithological grain size. Construct a high-dimensional engineering feature pool that includes the original well logging curves, texture feature curves, and gradient feature curves.
4. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 3, characterized in that, A hybrid filtering-wrapping strategy based on the binary sparrow search algorithm is employed to screen optimal feature combinations sensitive to lithological grain size identification from a high-dimensional engineering feature pool, including: Preliminary selection by filtering method: Based on the maximum information coefficient, a preliminary selection set of features is obtained from the high-dimensional engineering feature pool. Based on the variance inflation factor test, a final selection set of features is obtained from the preliminary selection set by filtering method. Wrap-up selection: The features selected by the filtering method are encoded into binary feature vectors. The classification accuracy of the LSSVM model is used as the first fitness function. The binary sparrow search algorithm is used to select the best combination of attribute features from the feature set selected by the filtering method.
5. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 4, characterized in that, Based on the maximum information coefficient, a preliminary set of features is obtained from the high-dimensional engineering feature pool, including: The maximum information coefficient algorithm is used to calculate the nonlinear correlation strength between each feature and rock grain size category in the high-dimensional engineering feature pool. Based on the nonlinear correlation strength, the nonlinear correlation strength is selected. Features with MIC (Micro-Matching Detection) thresholds are selected and removed, while nonlinear correlation strengths are retained. The features of the MIC screening threshold are used as preliminary screening features to obtain the preliminary screening feature set; The final filter feature set is obtained from the initial screening features based on the variance inflation factor test, including: For the initial screening feature set, a multiple linear regression is performed with the k-th initial screening feature as the dependent variable and the remaining initial screening features as independent variables, and the coefficient of determination of the multiple linear regression is calculated. Based on the coefficient of determination Calculate the variance inflation factor of the k-th preliminary screening feature: Iteratively calculate the variance inflation factor for each initially screened feature, and remove those with variance inflation factors. The VIF screening threshold is used to select features until the variance inflation factor of all initially screened features is less than the VIF screening threshold, thus obtaining the filter-selected feature set.
6. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 5, characterized in that, The binary sparrow search algorithm is used to select the optimal combination of attribute features from the feature set selected by the filtering method. The process includes: constructing a fitness function based on the classification accuracy of LSSVM: dividing the feature set selected by the filtering method into a training set and a validation set, training the LSSVM model using a K-fold cross-validation strategy, using the average classification accuracy of the LSSVM model on the validation set as the first fitness function, using the binary feature vectors of the filtered features as optimization variables, and using the binary sparrow search algorithm for iterative search to obtain the binary feature vectors representing the selected features. The filtered feature combination represented by the binary feature vectors representing the selected features after the iteration is taken as the optimal feature attribute combination.
7. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor construction according to any one of claims 1-6, characterized in that, The lithological grain size factor model constructed based on the optimal feature combination is as follows: in, For depth point Lithological grain size factor For the first The best attribute features in depth The value after normalization; For the first The weights of each feature, This represents the optimal number of attribute features.
8. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 7, characterized in that, The optimal lithological grain size factor model is obtained by optimizing the weights of the lithological grain size factor model using a sparrow search algorithm. This includes: using the weight vector and bias term of the lithological grain size factor model as optimization variables, maximizing the inter-class discrimination of coarse sand, medium sand, fine sand, silt, mudstone, and pebble sandstone as the optimization objective, introducing the generalized Fisher discrimination criterion as the second fitness function, and using the sparrow search algorithm to optimize the weight vector and bias term of the lithological grain size factor model to obtain the optimal lithological grain size factor model.
9. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor as described in claim 8, characterized in that, The sparrow search algorithm is used to optimize the weight vector and bias term of the lithological grain size factor model, including: Step 4.1: Using the optimal combination of attribute features as input, calculate the within-class scatter matrix. Inter-class scatter matrix According to the inner scatter matrix and inter-class scatter matrix Define a second fitness function based on the Fisher criterion: , Let be the weight vector to be optimized. It is the minimum value. Indicates transpose; Step 4.2: Map the weight vector and bias of the lithological grain size factor model to be optimized to the spatial location of individual sparrows, set the sparrow search algorithm parameters, and initialize the population; Step 4.3: Calculate the fitness value of each individual in the population using the second fitness function, select the discoverer and the newcomer according to the fitness value ranking, and record the global optimal position; Step 4.3: Update the positions of the discoverer and the joiner according to the position update formula; Step 4.4: Randomly select vigilant individuals from the discoverers and participants, and update their positions using the vigilant position update formula; Step 4.5: Repeat steps 4.3-4. When the maximum number of iterations is reached or the fitness value no longer increases for several consecutive generations, output the global optimal position.
10. The SSA-LSSVM reservoir lithology fine prediction method based on lithology grain size factor construction according to any one of claims 3-6, 8 or 9, characterized in that, The entire study area is predicted based on the trained SSA-LSSVM lithological grain size identification model. This includes: preprocessing the target logging curves to be predicted in the study area, calculating texture and gradient features using the S102 method to obtain texture and gradient feature curves, calculating the lithological grain size factor for each depth point using the optimal lithological grain size factor model of S104, generating continuous lithological grain size factor curves, fusing the lithological grain size factor curves with the target logging curves, texture and gradient feature curves as input, and identifying the specific lithological grain size level corresponding to each depth point using the trained SSA-LSSVM lithological grain size identification model.