A modeling method for reservoir prediction, a reservoir prediction method and a system

Abstract

The present invention discloses a modeling method, a reservoir prediction method, and a system for reservoir prediction. The method first prepares an observation sample dataset containing seismic attributes, then uses a clustering method to select variables and extract key variables to construct a working dataset. Through pseudo-labeling technology, limited labeled samples are expanded into fully labeled samples, and information from unlabeled data is used to optimize model training. The introduction of pseudo-labels significantly improves the recognition accuracy of categories with a small number of samples when labeled data is insufficient. The present invention is particularly suitable for practical scenarios such as oil and gas prediction, reservoir modeling, and well logging data analysis, and effectively solves the problem of reduced prediction accuracy due to label scarcity.

Description

Technical Field

[0001] The present invention relates to a modeling method for oil reservoir prediction, and also relates to a corresponding reservoir prediction method and a reservoir prediction system, belonging to the technical field of data processing. Background Art

[0002] Well logging data, a vertical depth profile acquired by logging tools, can reveal geological structural characteristics, such as lithology and reservoir type. This data is crucial for improving geological modeling and enhancing oil exploration efficiency. However, due to limitations such as poor wellbore conditions and high labor and material costs, labeled data typically only accounts for a small fraction of the total data. This scarcity of labeled data significantly reduces the effectiveness of supervised learning-based prediction models, making reservoir prediction a challenging task.

[0003] To address this challenge, semi-supervised learning (SML) has gained increasing attention in reservoir prediction in petroleum geological exploration. The core of semi-supervised learning lies in combining two types of data: (1) a small or medium-sized "labeled" dataset containing observations of the outcome variable and a set of covariates; and (2) a much larger "unlabeled" dataset containing only observations of the covariates. This property makes semi-supervised learning particularly suitable for big data applications where the outcome variable is difficult to obtain but the covariates are easily accessible.

[0004] In their 2021 study, Azriel et al. proposed a semi-supervised regression (SSR) model designed to improve least squares estimation (LSE) by transforming the regression problem into a mean estimation problem. This model successfully demonstrated the potential of semi-supervised regression models for oil reservoir prediction and provided an effective solution to the problem of scarce labeled data. The key to semi-supervised regression models is to leverage information from unlabeled samples to provide additional optimization support for the model, thereby achieving asymptotically uniform improvements in least squares estimation. Specifically, by decomposing the regression problem into multiple independent mean estimation problems, the parameter estimation process is simplified, while the predictive performance of the semi-supervised regression model is further optimized by leveraging unmodeled nonlinear features that may exist in the data. Summary of the Invention

[0005] The primary technical problem to be solved by the present invention is to provide a modeling method for reservoir prediction to realize a semi-supervised regression model with pseudo labels.

[0006] Another technical problem to be solved by the present invention is to provide a reservoir prediction method based on the semi-supervised regression model with pseudo labels.

[0007] Another technical problem to be solved by the present invention is to provide a corresponding reservoir prediction system.

[0008] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] According to a first aspect of an embodiment of the present invention, a modeling method for reservoir prediction is provided, comprising the following steps:

[0010] Step 1: Prepare a data set, which contains Y observation samples, each of which consists of N seismic attributes, which are numerical features that characterize logging events;

[0011] Step 2: Use clustering method to select variables for the data set and screen out P key variables;

[0012] Step 3: Based on the data set obtained in step 1 and the P key variables screened in step 2, construct a Y×P working data set;

[0013] Step 4: On the working dataset, a limited number of labeled samples are expanded into fully labeled samples by using pseudo-labeling technology to train the semi-supervised regression model, and a semi-supervised regression model with pseudo-labels is generated based on the training results of the PI estimator;

[0014] Among them, Y, N, and P are all positive integers.

[0015] According to a second aspect of an embodiment of the present invention, a reservoir prediction method is provided, comprising the following steps:

[0016] Input the dataset to be predicted into the semi-supervised regression model with pseudo labels constructed by the above modeling method;

[0017] A semi-supervised regression model with pseudo labels is used to predict reservoirs and output the prediction results.

[0018] According to a third aspect of an embodiment of the present invention, a reservoir prediction system is provided, comprising a processor and a memory, wherein the processor and the memory are coupled; wherein the memory is used to store a computer program; and the processor runs the computer program stored in the memory to implement the above-mentioned reservoir prediction method.

[0019] Compared with the existing technology, the present invention effectively utilizes unlabeled data through pseudo-labeling technology, expands labeled samples, and significantly improves the accuracy of reservoir prediction. In the case of scarce labeled data, the present invention shows higher prediction effectiveness, especially improving the recognition accuracy of categories with a small number of samples. Experimental results show that the present invention is superior to other methods in multiple evaluation indicators, including conventional semi-supervised regression models and other classical regression methods. Among them, the introduction of pseudo-labels significantly improves the prediction performance of the model in all cases, especially when labeled data is insufficient, the recognition ability of rare geological features is particularly prominent. Therefore, the present invention effectively solves the problem of reduced prediction accuracy due to scarcity of labels, significantly improves the accuracy of reservoir modeling, oil and gas prediction and logging data analysis, and is particularly suitable for actual scenarios in geological applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Histogram of simulated data variables used for actual reservoir prediction;

[0021] Figure 2 for Figure 1 Correlation matrix plot of the simulated data shown;

[0022] Figure 3A A schematic diagram of a flow chart of a modeling method provided in an embodiment of the present invention;

[0023] Figure 3B for Figure 3A Schematic diagram of the pseudo-labeling technology used in;

[0024] Figure 4 This is a schematic diagram of seven variables obtained through variable selection in the first embodiment of the present invention;

[0025] Figure 5 Schematic diagram of original seismic layers and well locations in real data;

[0026] Figure 6 This is a schematic diagram of the river channel on the seismic profile in the real data;

[0027] Figure 7 This is a working data set variable diagram obtained by performing variable selection based on real data in the first embodiment of the present invention;

[0028] Figure 8 A scatter plot comparison of prediction results of Well No. 1 in a real well logging dataset using the modeling method provided by an embodiment of the present invention;

[0029] Figure 9 A reservoir property heat map of the prediction results obtained by using the modeling method provided by the embodiment of the present invention;

[0030] Figure 10This is a schematic diagram of 23 variables obtained through R-type clustering in the second embodiment of the present invention;

[0031] Figure 11 In the second embodiment of the present invention, Figure 10 Schematic diagram of the six variables selected after clustering using principal component analysis. DETAILED DESCRIPTION

[0032] The technical content of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] The technical concept behind the embodiments of this invention is to expand limited labeled samples into fully labeled samples using pseudo-labeling techniques, fully utilizing the information from unlabeled data to improve the accuracy and robustness of reservoir prediction. The specific process includes preparing an observation sample dataset containing seismic attributes, using clustering methods to select variables to extract key variables, constructing a working dataset, and expanding the labeled samples using pseudo-labeling techniques. Finally, a semi-supervised regression (SSR) model is trained to obtain an optimized reservoir prediction model. This method significantly improves the accuracy of reservoir prediction when labeled data is scarce. It is particularly suitable for categories with small sample sizes and enhances the ability to identify rare geological features.

[0034] Here, we first briefly describe the semi-supervised regression model itself. It is a method for performing regression analysis using limited labeled data and a large amount of unlabeled data. Considering only partial information, and assuming that the number of labeled samples, m, is at least proportional to the number of unlabeled samples, n, the semi-supervised regression model transforms the regression problem into a mean estimation problem for each parameter. Through a multiplication step, the semi-supervised regression model provides asymptotically better parameter estimates than the standard least squares estimation (LSE), thereby obtaining the optimal linear predictor. Specifically, the semi-supervised regression model exploits the unmodeled nonlinear information that may exist in the unlabeled data and optimizes the model parameters by adjusting the regression coefficients. For example, the p-dimensional linear regression model shown in the following equation simplifies the complex regression analysis process by decomposing the regression problem into p independent mean estimation problems, solving each parameter estimation problem separately. This asymptotically provides more accurate predictions.

[0035]

[0036] in, are the coefficients of the optimal linear predictor, defined as .

[0037] Remaining items satisfy .

[0038] Then, the p-dimensional estimation process is simplified into p independent simple regression problems, and the p mean estimation problems are solved separately.

[0039] To define the PI estimator, let:

[0040]

[0041] .

[0042] in, Based on the size of The empirical mean of the complete sample X of . Now define

[0043] as well as

[0044]

[0045] here, Indicates the entire The adjustment of the sample. Accordingly, the PI estimator is:

[0046]

[0047] in, represents the least squares estimate of the regression model.

[0048]

[0049] Here, use replace ,get:

[0050] .

[0051] in,

[0052]

[0053] On the other hand, the inventors used two sets of public domain datasets in their research. One set is the forward simulation data of a river channel, and the other set is the real data from a certain built-up area. The structure of the simulation data and the real data is as follows:

[0054] Simulated data: Contains 10,201 observation samples, each composed of 31 seismic attributes. Each seismic attribute is a numerical feature representing the physical properties of seismic waves, such as amplitude, frequency, and arrival time, which together characterize an earthquake event. The dataset can be represented as a 10,201×31 matrix, where each row represents a sample and each column represents an attribute. Labeled and unlabeled samples are mixed in the dataset.

[0055] The real data also contains 10,201 observations, each consisting of 31 seismic attributes. Due to insufficient exploration in this area, only 50 of these observations are labeled, leaving the remaining 10,151 unlabeled. The dataset structure is similar to the simulated data: a 10,201 × 31 matrix with a mixed distribution of labeled and unlabeled samples.

[0056] Through these two sets of data, the inventors verified the effectiveness of the embodiment of the present invention in the case of scarce labeled data. The structure of the simulated data is shown in Table 1 below.

[0057] Table 1 Simulation data

[0058]

[0059] Figure 1 The distribution of the covariates is shown, showing that most variables have peaks near the central value and are relatively symmetrical around the mean. No variable in the figure has a uniform distribution. Figure 1 and Figure 2 Histograms and correlation matrices for variables (X1–X31) are provided, comprehensively illustrating the characteristics of the variables and their interrelationships. The histograms reveal that variables X1–X6, X9–X12, X17–X18, X22–X26, and X28–X30 exhibit relative symmetry, while variables X3, X7–X8, X13–X16, X19–X21, and X31 exhibit skewed distributions. In particular, variables X10–X11 exhibit a multimodal distribution. Furthermore, the distribution ranges of the variables vary. For example, X1–X6 and related variables have a narrow distribution range, while X3 and its corresponding variables have a wide distribution range. Notably, variable X13 contains an outlier. The correlation matrix reveals intervariate correlations: for example, there is a strong positive correlation between X1 and X3 and subsequent pairs; there is a strong negative correlation between X1 and X29, and other pairs; and there is weak or no correlation between X1 and X13. Diagonal elements are all 1, indicating perfect correlation between the variables themselves. Based on an in-depth analysis of the characteristics of these variables, effective variable selection can be performed, retaining variables that are strongly correlated with the target variable or that contribute significantly to the overall data pattern, while eliminating weakly correlated or meaningless variables. By additionally utilizing unlabeled data and correcting the estimation of linear parameters, the prediction accuracy of the model can be improved even if there are unmodeled nonlinear relationships in the real model. Based on the above comprehensive understanding of the characteristics of reservoir prediction data, an embodiment of the present invention provides a semi-supervised regression model with pseudo labels, which aims to significantly improve the accuracy of reservoir prediction.

[0060] First embodiment

[0061] like Figure 3A and Figure 3BAs shown, a modeling method for reservoir prediction provided by the first embodiment of the present invention includes at least the following steps.

[0062] Step 1: Prepare the dataset.

[0063] This dataset contains Y observation samples, each composed of N seismic attributes. Each seismic attribute is a numerical feature representing some physical properties of the seismic wave, such as amplitude, frequency, and arrival time. These properties together characterize the logging event. The dataset can be represented as a Y × N matrix, where each row represents a sample and each column represents a seismic attribute.

[0064] Among these Y samples, M samples are labeled (hereinafter referred to as labeled samples), and the remaining Y-M samples are unlabeled (hereinafter referred to as unlabeled samples). Labeled samples and unlabeled samples are mixed in the dataset.

[0065] Step 2: Use Q-type clustering based on target value correlation for variable selection.

[0066] In scenarios such as feature selection, dimensionality reduction, and variable grouping in machine learning, R-type clustering is usually more commonly used. It can help screen out irrelevant features, thereby effectively reducing the complexity of the model. In contrast, Q-type clustering mainly indirectly evaluates the importance of variables by analyzing the similarities between samples. Although R-type clustering is good at grouping variables, when dealing with features with strong correlations, it may classify these related variables into the same category, but it cannot further eliminate redundant variables. In view of the variable selection requirements of the present invention, Q-type clustering becomes a more appropriate choice because it can more directly reflect the importance of variables.

[0067] Based on the above selections, step 2 includes the following sub-steps:

[0068] Step 2.1 Use the Pearson correlation formula to calculate the correlation between each variable and the target value to identify those variables that exhibit a strong linear relationship. The target value is the labeled well logging data, i.e., the Y value.

[0069] In one embodiment of the present invention, the formula for Pearson correlation is:

[0070]

[0071] Where x and y represent the individual data points of the input variables (such as seismic attributes) and target values ​​(such as well logging data) of M labeled samples, respectively. In the application scenario of reservoir prediction, is the earthquake attribute of the ith labeled sample, is the logging data of the ith labeled sample, and Here, a series of operations are performed using M labeled X and Y to perform variable selection.

[0072] Based on the calculation results, N1 (N1<N) highly correlated variables are identified.

[0073] Step 2.2: Use Q-type clustering to classify the variables into different groups based on the correlation patterns between the N1 variables (nine variables were selected in this example) identified in the previous step and the target value. The correlation pattern is the relationship structure between the target value and the N1 variables, primarily reflecting the similarities and differences between the variables. This is categorized using Q-type clustering.

[0074] The clustering objective can be expressed as minimizing the within-group variance , the formula is as follows:

[0075]

[0076] in, Indicates the The clustering results, is the center of the cluster, are the data points (variables) within each cluster.

[0077] According to the clustering results, further screening is performed from N1 variables, and finally P (P < N) key variables are screened out (P = 7 in this embodiment), such as Figure 4 shown.

[0078] Step 3: Use the data set from step 1 and the P key variables selected in step 2 to construct a working data set. The working data set is a Y×P matrix.

[0079] Step 4: On the working dataset, pseudo-labeling techniques are used to expand the limited number of labeled samples to fully labeled samples. Pseudo-labels are generated by screening unlabeled samples with high consistency, and a fully labeled training set is constructed for training the semi-supervised regression model. Subsequently, based on the training results of the PI estimator, a semi-supervised regression model with pseudo-labels is generated.

[0080] In the following description, the definitions of various parameters or symbols can refer to the definitions in the semi-supervised regression model introduced above.

[0081] In one embodiment of the present invention, the method specifically includes the following sub-steps:

[0082] 4-1) Set the input and output as follows:

[0083] enter:

[0084] Observation vector ;

[0085] Matrix of working dataset (with p columns) ;

[0086] Maximum number of iterations: ;

[0087] Consistency threshold: ;

[0088] Number of self-trained models: ;

[0089] Output:

[0090] For each k=1,...p PI estimator ;

[0091] 4.2) Initialize the semi-supervised regression model. Set the iteration counter And satisfy Remove the kth column from the matrix X and add an intercept column to X .

[0092] 4.3) Calculate the regression coefficient. Use labeled samples to solve:

[0093] ,

[0094] Calculate the unlabeled sample i component .

[0095] 4.4) Adjusting the response variable .

[0096] 4.5) Standardized Regressors .

[0097] 4.6) Adjusting the Regressor

[0098] .

[0099] 4.7) Calculate the PI estimator .

[0100] 4.8) Calculate the consistency score.

[0101] 4.9) Determine consistency indicators.

[0102] 4.10) Update the data and retrain the model.

[0103] 4.11) Return results .

[0104] It should be noted that the labeled samples in step 4 are obtained through pseudo-labeling technology. Pseudo-labeling is a simple but efficient method that aims to improve the performance of the predictor by simultaneously utilizing labeled data and unlabeled data. In one embodiment of the present invention, the pseudo-labeling technology includes the following sub-steps: training an initial semi-supervised regression model based on labeled samples, predicting unlabeled samples, and generating prediction results of multiple models; calculating the consistency score of each unlabeled sample, and screening out unlabeled samples below the set consistency threshold as pseudo-labels; merging the pseudo-labels with the original labeled samples, updating the training set and iteratively optimizing the model until the convergence conditions are met.

[0105] Specifically, the model is first trained only on the labeled dataset. In each iteration, the model uses the prediction results of the previous iteration as the predicted values ​​of the unlabeled samples, treating them as the true labels. In order to improve the robustness of the model, this process is repeated n times (n is the number of models), and different random seeds are used for each training to generate diverse prediction results. After obtaining prediction results from multiple models, the next step is to evaluate the consistency of the pseudo-labels. The prediction results of all models are stored in a list, and by calculating the average of these prediction results, the prediction stability of each unlabeled sample can be evaluated. Specifically, the consistency evaluation is achieved by calculating the consistency score of each unlabeled sample in the predictions of multiple models, thereby screening out those samples with more stable prediction results. These pseudo-labels that have been verified to be consistent will be used to expand the labeled dataset and further optimize the training effect of the model.

[0106]

[0107] in, is the prediction result of the kth model for the unlabeled sample. Specifically, the consistency score Ci of the i-th unlabeled sample (unlabeled sample i) is calculated, which is the sum of the absolute differences between the prediction value of each model and the mean.

[0108]

[0109] in, is the prediction of the kth model for the unlabeled sample i. This consistency check is designed to ensure that only those predictions that are consistent across different models are considered as pseudo labels.

[0110] Based on the set consistency threshold ( ) and the number of self-trained models , select consistency scores lower than To ensure the reliability of the selected pseudo labels:

[0111] .

[0112] These pseudo-labels are merged with the original labeled samples Y to form an updated labeled dataset, enabling the model to better utilize unlabeled data. In the process of merging pseudo-labels, the feature matrix X of the working dataset is also expanded to include data from both the original labeled samples and the added pseudo-labels. This allows the model to leverage a richer dataset for training.

[0113] In one embodiment of the present invention, pseudo-labeling techniques are used to obtain labeled samples, including the following sub-steps: First, input is an observation vector Y (original labeled samples), a feature matrix X (containing p columns), a maximum number of iterations, a consistency threshold, and the number of self-trained models. In each iteration, an empty list (predictions_list) is initialized to store the prediction results. Next, each model is trained by setting a random seed, training the model using the original labeled samples Y, and making predictions for unlabeled samples, storing the prediction results in a list. Next, the average prediction value for each unlabeled sample is calculated, and a consistency score is calculated based on the predictions of all models. The consistency score assesses the stability of the predictions by calculating the sum of the absolute differences between each model's prediction and the average prediction value. Based on the consistency score and the consistency threshold, samples with high stability are selected as pseudo-labels. Finally, these pseudo-labels are merged with the original labeled samples Y to form a new labeled dataset for subsequent model training. This process, through multiple iterations and multi-model predictions, combined with consistency checks, effectively utilizes unlabeled data and enhances the robustness and predictive performance of the model.

[0114] The algorithm has two stopping criteria: First, when the maximum absolute change in the model parameters is less than a set consistency threshold, indicating that the model parameters have stabilized and have not changed significantly, the iteration process terminates; second, if the number of iterations reaches a pre-set maximum allowable value, the iteration also stops. This means that as long as the model converges before the maximum number of iterations is reached, the iteration can be terminated early. Pseudo-labeling technology drives the self-training process, improving model performance by training on both labeled and unlabeled data. Specifically, for each unlabeled sample, predictions are first made using multiple independently trained models, and the average of their predictions is calculated as a candidate pseudo-label. A consistency score is then calculated by evaluating the dispersion of these predictions. Ultimately, only unlabeled samples with a consistency score below the consistency threshold are retained as pseudo-labels and included in the training set. This mechanism ensures the reliability and accuracy of pseudo-labels by strictly screening samples with highly consistent predictions.

[0115] To verify the practical effect of the present invention, we used a real well logging data set for experimentation. The data set contains 396,066 well logging data points distributed in a working grid and covering 9 different geological attributes of the formation.

[0116] Among these well logging data, 22 wells have labeled values, while the remaining 396,044 wells have no labeled values. Therefore, the composition of the entire dataset is: the number of unlabeled samples is 396,044, the number of labeled samples is 22, and each sample contains 9 feature variables.

[0117] In the experiment, data from 16 wells were selected from the 22 known wells for model training, while the data from the remaining 6 wells were used as a blind well test set to evaluate the model's prediction performance. The results of the blind well test are shown in Table 2.

[0118] By comparing the prediction results of the pseudo-labeled semi-supervised regression (PL-SSR) model provided by the embodiment of the present invention and the conventional semi-supervised regression (SSR) model, the performance and advantages of the PL-SSR model on real datasets can be evaluated.

[0119] Table 2 Blind well test results

[0120]

[0121] Figure 5 The original seismic layers and well locations of the real well logging dataset are shown. Figure 6 The river channel position of the dataset on the seismic profile is shown. The variables of the working dataset are selected by the Q-type clustering method. Figure 7 Selected variables are presented, including: peak amplitude (PeakAmplitude), spectral filters at different frequencies (20Hz, 30Hz, 40Hz, 50Hz, 100Hz) (SF_amplitude_20, SF_amplitude_30, SF_amplitude_40, SF_amplitude_50, SF_amplitude_100), seismic impedance characteristics (standard Prg_CIP), standardized attributes (standardperigram), standardized porosity calculation parameters (standard_Prg_CIP), and standardized reflection intensity (standardRSt).

[0122] The specific meaning of each variable is as follows:

[0123] Peak Amplitude: Related to the reflection intensity of the formation, high values ​​may indicate hard or compact rock formations, and low values ​​may be associated with loose sedimentary layers.

[0124] SF_amplitude_20~100: The amplitudes of different frequencies reflect geological features at different depths or scales. High frequencies (such as 100 Hz) reveal details or cracks, while low frequencies (such as 20 Hz) reflect macroscopic sedimentary features.

[0125] Standard Prg_CIP: Standardized seismic impedance or key point data used to analyze reflection contrast between geological layers and identify underground fluids and lithology.

[0126] Standard RSt: Standardized reflection intensity, used for reservoir identification and revealing underground lithology or fluid distribution.

[0127] Standard perigram: Standardized seismic profile data used to identify seismic reflection boundaries, faults, sedimentary environments, etc.

[0128] standard_Prg_CIP: Standardized porosity calculation method or parameters.

[0129] Figure 8 The prediction results of the PL-SSR model provided by the first embodiment of the present invention and the conventional SSR model are compared. Figure 8 It can be seen that the prediction results of the PL-SSR model are closer to the true value, indicating that its prediction effect is better than that of the conventional SSR model. Figure 9 The reservoir property heat map of the PL-SSR model prediction results is displayed, which clearly presents the characteristics and direction of the river channel. The river channel outline is very distinct, further confirming the effectiveness of the pseudo-labeling technology and highlighting the significant advantages of the PL-SSR model over the conventional SSR model.

[0130] Second embodiment

[0131] The second embodiment of the present invention discloses another modeling method for reservoir prediction. Its main steps are similar to those of the first embodiment, but it employs a different combination of techniques in the variable selection method. Specifically, in step 2, the second embodiment combines R-type clustering and principal component analysis (PCA) for variable selection and dimensionality reduction.

[0132] In step 2.1, R-type clustering is first performed. R-type clustering is usually based on a distance metric, such as Euclidean distance, to evaluate the similarity between data points.

[0133] The distance is defined as:

[0134]

[0135] Here, x represents the seismic attributes of the data point, y represents the well log data of the data point, and n is the number of features or dimensions.

[0136] The distance definition here involves the seismic attributes and well logging data of the data points, where the number of features or dimensions is a key parameter. Through R-type clustering, N' variables (N' is less than N) can be selected from the original N variables. In this embodiment, this number is determined to be 23 variables, such as Figure 10 shown.

[0137] Next, perform principal component analysis (PCA) in step 2.2. The goal of PCA is to reduce the dimensionality of the data while preserving the maximum variance in the data. This is achieved by calculating the eigenvalues ​​associated with each principal component, where the magnitude of the eigenvalue reflects the amount of information contained in the corresponding principal component.

[0138] Principal component analysis can be expressed as:

[0139]

[0140] in, represents the eigenvalue associated with the i-th principal component, and k represents the number of principal components required to retain at least 90% of the total variance.

[0141] In this example, the number of principal components was selected based on the criterion of retaining at least 90% of the total variance. This approach ensures that while reducing the data dimensionality, it also retains as much of the key information as possible. In this way, principal component analysis helps further optimize the variable selection process, ensuring that the final selected variables are more representative of the main characteristics of the data.

[0142] After R-type clustering, N' variables (N'<N) are screened out from the original N variables. This number is determined to be 23 variables in this embodiment. Next, principal component analysis is performed on these 23 variables. Principal component analysis is a statistical method used to reduce the dimension of data while retaining the maximum variance in the data. By calculating the eigenvalue of each principal component, its contribution to the total variance of the data can be determined. In this embodiment, the number of principal components is selected based on the criterion of retaining at least 90% of the total variance. Based on the results of the principal component analysis, P key variables (P<N') are further screened out from these 23 variables. These variables reduce the complexity of the data while retaining key information, thereby optimizing the variable selection process and providing a more efficient data set for subsequent modeling.

[0143] To comprehensively evaluate the performance of the PL-SSR model, the inventors systematically compared it with classical regression methods, including linear regression, ridge regression, lasso regression, and random forest regression. This comparison covered both the first embodiment (Method 1) and the second embodiment (Method 2), and detailed analysis was performed on the prediction results of each model under three different conditions. The experimental results revealed three key conclusions:

[0144] (1) The effectiveness of reservoir prediction is highly sensitive to the number of available samples, especially when labeled data is scarce, the model performance varies significantly.

[0145] (2) After the introduction of pseudo-labeling technology, the overall performance of the semi-supervised regression model was significantly improved, proving the effectiveness of pseudo-labeling technology in enhancing the model's predictive ability.

[0146] (3) Under the framework of the PL-SSR model, the model not only effectively addresses the challenge of insufficient labeled data, but also particularly improves the recognition accuracy of categories with a small number of samples, which is particularly important for the recognition of rare categories in geological applications.

[0147] Specifically, the PL-SSR model demonstrated superior predictive performance under all test conditions, with low root mean square error (RMSE) and mean absolute error (MAE), and a high coefficient of determination (R²), indicating that the model can more accurately predict reservoirs. In particular, when labeled data is scarce, the PL-SSR model significantly improves the recognition accuracy of classes with fewer samples compared to other methods. This has important implications for practical applications such as oil and gas prediction, reservoir modeling, and well logging data analysis.

[0148] Table 3 Data simulation results

[0149]

[0150] Table 3 shows the prediction performance of different models using the first embodiment (method 1) and the second embodiment (method 2) on the original data and the data processed by the two different methods. The evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

[0151] When comparing the SSR, PL-SSR, Lasso, Ridge, RF, and PL-RF methods, Method 1 performed better in most cases, with lower RMSE and MAE values ​​and higher R² values, indicating that Method 1 generally has higher prediction accuracy. However, under the PL-Lasso method, Method 2 exhibited lower RMSE and MAE values ​​and a slightly higher R² value, indicating that Method 2 has better prediction performance in specific situations.

[0152] This concludes that Method 1 outperforms Method 2 in most cases. Notably, the introduction of pseudo-labels improves model performance in all cases, highlighting the effectiveness of pseudo-labeling in enhancing model performance. Overall, the semi-supervised regression model exhibits low RMSE and MAE values ​​and high R² values. The addition of pseudo-labels further reduces RMSE and improves model performance.

[0153] In summary, when applied to well logging data, the present invention draws the following conclusions by comparing the overall performance of different variable selection techniques with common regression methods:

[0154] Overall Performance: Both PL-SSR and SSR show superior overall performance, with lower RMSE and MAE and higher R² values, indicating that their regression predictions are more accurate.

[0155] Impact of pseudo-labels: Adding pseudo-label training significantly improves the overall performance of the model. Especially for SSR, the regression prediction accuracy is improved most significantly after adding pseudo-labels.

[0156] Therefore, the present invention successfully solves the problem of reduced prediction accuracy caused by label scarcity and significantly improves prediction accuracy. It is particularly suitable for geological applications such as oil and gas prediction, reservoir modeling, and logging data analysis.

[0157] Third embodiment

[0158] Based on the first or second embodiment, the third embodiment of the present invention further provides a reservoir prediction method, comprising the following steps:

[0159] The dataset to be predicted is input into a semi-supervised regression model with pseudo labels, where each earthquake attribute in the dataset is a numerical feature representing the physical properties of the seismic wave;

[0160] Use a semi-supervised regression model with pseudo labels to predict reservoirs and output prediction results;

[0161] The pseudo-labeled semi-supervised regression model is constructed by the modeling method for reservoir prediction described in the first embodiment (method 1) or the second embodiment (method 2).

[0162] Fourth embodiment

[0163] like Figure 11 As shown, based on the above-mentioned reservoir prediction method, a fourth embodiment of the present invention further provides a reservoir prediction system. The reservoir prediction system includes one or more processors and a memory. The memory is coupled to the processor and is configured to store one or more programs. When the programs are executed by the processor, the processor implements the reservoir prediction method described in the above-mentioned embodiment.

[0164] The processor is used to control the overall operation of the reservoir prediction system to complete all or part of the steps of the above-mentioned reservoir prediction method. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. The memory is used to store various types of data to support the operation of the reservoir prediction system. This data may include, for example, instructions for any application or method operating on the reservoir prediction system, as well as application-related data. The memory can be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, etc.

[0165] In another exemplary embodiment, the present invention further provides a computer-readable storage medium comprising program instructions. When executed by a processor, the program instructions implement the steps of the reservoir prediction method described in any of the aforementioned embodiments. For example, the computer-readable storage medium may be the aforementioned memory comprising the program instructions, which may be executed by a processor to perform the aforementioned reservoir prediction method and achieve the same technical effects as the aforementioned method.

[0166] It should be noted that the above embodiments are merely examples, and the technical solutions of the various embodiments may be combined, and the order of the steps may be changed, all within the scope of protection of the present invention.

[0167] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature identified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0168] The above describes in detail the modeling method, reservoir prediction method, and system for reservoir prediction provided by the present invention. Any obvious modification made thereto by one skilled in the art without departing from the essence of the present invention would constitute an infringement of the patent rights of the present invention and would result in corresponding legal liability.

Claims

1. A modeling method for reservoir prediction, characterized in that The steps include: Step 1: Prepare a data set, which contains Y observation samples, each of which consists of N seismic attributes, which are numerical features that characterize logging events; Step 2: Use clustering method to select variables for the data set and screen out P key variables; Step 3: Based on the data set obtained in step 1 and the P key variables screened in step 2, construct a Y×P working data set; Step 4: On the working dataset, a limited number of labeled samples are expanded into fully labeled samples using pseudo-labeling technology for training the semi-supervised regression model, and a semi-supervised regression model with pseudo-labels is generated based on the training results of the PI estimator; where Y, N, and P are all positive integers; The pseudo-labeling technique in step 4 includes the following sub-steps: An initial semi-supervised regression model is trained based on labeled samples, and predictions are made for each unlabeled sample to generate prediction results for multiple models. Calculate the consistency score of each unlabeled sample and select the unlabeled samples below the set consistency threshold as pseudo labels; Merge the pseudo labels with the original labeled samples, update the training set and iteratively optimize the model until the convergence condition is met; Among them, the consistency score C of the i-th unlabeled sample is i , is the sum of the absolute differences between each model's prediction and the mean of the predictions.

2. The modeling method for reservoir prediction according to claim 1, characterized in that The clustering method in step 2 is Q-type clustering, which specifically includes the following sub-steps: Using the Pearson correlation calculation formula, the correlation between each variable and the target value is calculated, and N1 highly correlated variables are screened out, wherein the target value is the labeled well logging data; The N1 variables are subjected to Q-type clustering. Based on the principle of minimizing the within-group variance, the variables are grouped and P key variables are selected.

3. The modeling method for reservoir prediction according to claim 1, characterized in that The clustering method in step 2 is a combination of R-type clustering and principal component analysis, which specifically includes the following sub-steps: N' variables are selected from the original variables through R-type clustering; A principal component analysis was performed on the N' variables, and P key variables were selected based on the criterion of retaining at least 90% of the total variance.

4. The modeling method for reservoir prediction according to claim 3, characterized in that The following sub-steps are included: For each unlabeled sample, we first use multiple independently trained models to make predictions, and calculate the average of the prediction results as candidate pseudo labels; then, we calculate the consistency score by evaluating the degree of dispersion of the prediction results; finally, we only retain unlabeled samples with a consistency score lower than the consistency threshold as pseudo labels and include them in the training set.

5. The modeling method for reservoir prediction according to claim 4, characterized in that The prediction stability of each unlabeled sample is evaluated using the following formula: ； in, is the prediction result of the kth model for unlabeled samples, where k is a positive integer.

6. The modeling method for reservoir prediction according to claim 5, characterized in that The consistency score Calculated by the following formula: ； in, is the prediction result of the k-th model for the i-th unlabeled sample, where i is a positive integer.

7. A reservoir prediction method, characterized in that The following steps are involved: Inputting the dataset to be predicted into the semi-supervised regression model with pseudo labels constructed by the modeling method according to any one of claims 1 to 6; A semi-supervised regression model with pseudo labels is used to predict reservoirs and output the prediction results.

8. A reservoir prediction system, characterized in that The system comprises a processor and a memory, wherein the processor and the memory are coupled; wherein the memory is used to store a computer program; and the processor runs the computer program stored in the memory to implement the reservoir prediction method according to claim 7.

Images