Reservoir thickness prediction method based on improved Stacking ensemble learning
Through the improved Stacking integrated learning method, combined with multiple machine learning algorithms and Bayesian optimization, the problems of insufficient generalization capability and information loss in reservoir thickness prediction are solved, and more efficient reservoir thickness prediction is achieved.
Patent Information
- Application Number
- CN202510650892.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-20
- Publication Date
- 2025-08-08
AI Technical Summary
In the prior art, a single machine learning algorithm has insufficient generalization capability in reservoir thickness prediction, and traditional Stacking integrated learning models may lose original data information, and the hyperparameter optimization efficiency is low.
The improved Stacking ensemble learning method is adopted, combining the random forest model, extreme gradient lifting algorithm and long-term and short time difference network as the basis learner, and the reservoir thickness prediction is predicted using a multivariate linear regression model, the prediction results of the basis learner and the meta learner are integrated, and the seismic attributes are optimized through Pearson correlation analysis.
The model generalization performance and prediction accuracy of reservoir thickness prediction are improved, the efficiency of hyperparameter optimization is improved, and the problems of insufficient generalization capabilities and information loss of a single model are solved.
Smart Images
Figure CN120447038A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of petroleum geophysical exploration, and in particular relates to a reservoir thickness prediction method based on improved Stacking ensemble learning. Background Art
[0002] Seismic attributes effectively reflect subsurface geological information and play a crucial role in all stages of oil and gas exploration and development. They have been widely applied in reservoir prediction. Seismic multi-attribute fusion is a key research area in seismic attribute analysis technology and has achieved promising results in reservoir prediction. Numerous methods exist, broadly categorized as linear model fusion, nonlinear model fusion, and color model fusion.
[0003] With the rapid development of artificial intelligence (AI) technology, machine learning methods are increasingly being used for seismic multi-attribute fusion. These methods primarily learn prior information about wellpoint reservoirs, establish a nonlinear mapping relationship between well and seismic data, and then use machine learning methods to quantitatively predict reservoirs. Common seismic attribute fusion methods are primarily based on single-class machine learning algorithms and lack the integration of multiple-class algorithms. Due to the complex and highly heterogeneous structure of underground reservoirs, different fusion algorithms are suitable for different oilfields, and even for different locations or intervals within the same oilfield. Therefore, exploring attribute fusion methods that integrate multiple algorithms to improve the generalization capability of seismic attribute fusion methods is of great significance. Summary of the Invention
[0004] The present invention aims to provide a method for predicting turbidite reservoir thickness based on improved stacking ensemble learning. The present invention specifically comprises the following steps:
[0005] Step 1: Extract multiple seismic attributes of the study layer and calculate the reservoir thickness of the well points. After processing and attribute optimization, combine them into sample data, and divide the sample data into an initial training set and an initial test set;
[0006] Step 2: Input the initial training set into the Stacking ensemble learning base learner for five-fold cross-validation and hyperparameter optimization. The prediction results are combined with the initial training set to obtain the meta-learner training set. At the same time, the initial test set is used for prediction and performance evaluation. The prediction results are combined to obtain the meta-learner test set.
[0007] Step 3: Input the meta-learner training set into the Stacking ensemble learning meta-learner for hyperparameter optimization to obtain a trained meta-learner. Input the multi-attribute data into the trained meta-learner for prediction to obtain the final reservoir thickness prediction result. Use the meta-learner test set to evaluate the performance of the trained meta-learner.
[0008] The multiple seismic attributes described in step 1 include bandwidth, number of zero crossings, arc length, root mean square amplitude, energy half time, average energy, average amplitude, dominant frequency, positive amplitude and.
[0009] The attribute optimization method described in step 1 is Pearson correlation analysis:
[0010]
[0011] Where x and y represent two random variables. represents the sample mean, and r represents the Pearson correlation coefficient.
[0012] The sample data described in step 1 is divided into a training set and a test set in a ratio of 7:3.
[0013] The stacking ensemble learning base learners described in step 2 include a random forest model, an extreme gradient boosting algorithm, and a long and short time difference network.
[0014] The five-fold cross-validation steps described in step 2 are:
[0015] The training set was divided into five separate subsets of approximately equal size and non-overlapping data.
[0016] Among the five subsets, one of them is selected in sequence as the validation set, and the remaining four subsets are used as training sets;
[0017] The model obtained through the training set is verified on the validation set to obtain the performance indicators of the model on the current validation set;
[0018] Repeat the steps until all subsets are used as validation sets, and obtain the performance indicators of the model after five training cycles;
[0019] The performance indicators of the model are averaged over five training runs to obtain the performance indicators of a single base learner.
[0020] The meta-learner training set and meta-learner test set described in step 2 are combined as follows:
[0021] During the five-fold cross-validation process, the prediction results of all base learners on the validation set are combined and merged with the initial training set as the training set of the meta-learner;
[0022] The test set of the meta-learner is the combination of the average prediction results of all base learners trained in the five-fold cross-validation process on the test set;
[0023] The performance indicators mentioned in steps 2 and 3 are mean square error and coefficient of determination:
[0024]
[0025] Where: yi represents the actual value of the i-th sample, represents the predicted value of the i-th sample, M represents the number of samples, y mean Represents the mean of the actual values of the sample set.
[0026] The hyperparameter optimization method described in steps 2 and 3 is Bayesian optimization.
[0027] The stacking ensemble learning meta-learner described in step 3 is a multiple linear regression model.
[0028] Through the above-mentioned design scheme, the present invention can bring the following beneficial effects:
[0029] This invention addresses the common shortcomings of conventional machine learning algorithms for reservoir thickness prediction and proposes a reservoir thickness prediction method based on improved Stacking ensemble learning. Stacking ensemble learning addresses the poor generalization capabilities of a single model. By incorporating the initial training set into the meta-learner training set during model fusion, it addresses the potential loss of some original dataset information in traditional Stacking ensemble learning models. Furthermore, by optimizing hyperparameters using a Bayesian algorithm, it addresses the computational inefficiency of traditional grid search. Consequently, this invention offers the advantages of excellent model generalization, high prediction accuracy, and efficient hyperparameter optimization. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 Flowchart of the present invention.
[0031] Figure 2 Pearson correlation analysis plot for seismic attribute optimization. DETAILED DESCRIPTION
[0032] The present invention is described in detail below with reference to the accompanying drawings and specific implementation steps.
[0033] The purpose of the present invention is to provide a turbidite reservoir thickness prediction method based on improved Stacking ensemble learning.
[0034] The present invention specifically comprises the following steps:
[0035] S11. Extract various seismic attributes of the study interval based on the interpreted horizon, including bandwidth, number of zero crossings, arc length, root mean square amplitude, energy half time, average energy, average amplitude, dominant frequency, positive amplitude, and;
[0036] S12. Calculate the reservoir thickness at the well point based on the interpreted horizon, SP logging curve and logging data;
[0037] S13, combining the well point attribute data as characteristic values and the well point reservoir thickness as target values into original sample data;
[0038] S14. Perform Pearson correlation analysis on the initial sample data to select seismic attributes that have good correlation with reservoir thickness:
[0039]
[0040] Where x and y represent two random variables. represents the sample mean, and r represents the Pearson correlation coefficient.
[0041] S15, combining the preferred seismic attributes as feature values and the well point reservoir thickness as target values to form initial sample data, and dividing it into an initial training set F and an initial test set T in a ratio of 7:3;
[0042] S21, divide the initial training set F into 5 equal parts F1, F2, F3, F4, and F5 according to the equal division principle;
[0043] S22. The base learners of Stacking ensemble learning include random forest model, extreme gradient boosting algorithm, and long short time difference network. Each base learner needs to perform the following operations. Taking the random forest model as an example, select one of F1, F2, F3, F4, and F5 as the validation set, and the remaining four as the training set. The random forest model is trained and the Bayesian algorithm is used to optimize the hyperparameters. The prediction sets a11, a12, a13, a14, and a15 of the validation set will be obtained in turn, and merged into A1 in order. The trained random forest model is used to predict the test set T, and b11, b12, b13, b14, and b15 are obtained respectively, and the average of them is B1.
[0044] S23. When the three base learners, random forest model, extreme gradient boosting algorithm and long short time difference network, are all trained, the predicted values A1, A2 and A3 of these three base learners on the training set will be obtained. After merging them horizontally with the initial training set F, they will be used as the training set F_meta of the meta-learner.
[0045] S24. The prediction values B1, B2, and B3 obtained by the three base learners, random forest model, extreme gradient boosting algorithm, and long short time difference network, on the initial test set T are also merged horizontally as the test set T_meta of the meta-learner.
[0046] S31. Input the meta-learner training set F_meta into the Stacking ensemble learning meta-learner multivariate linear regression model for training and Bayesian algorithm optimization of hyperparameters to obtain a trained meta-learner. Use the meta-learner test set T_meta to evaluate the performance of the trained multivariate linear regression model, using root mean square error and coefficient of determination as performance evaluation indicators:
[0047]
[0048] Where: y i represents the actual value of the i-th sample, represents the predicted value of the i-th sample, M represents the number of samples, y mean Represents the mean of the actual values of the sample set.
[0049] S32. Input the multi-attribute data into the trained multivariate linear regression model for prediction to obtain the final reservoir thickness prediction result.
Claims
1. A reservoir thickness prediction method based on improved Stacking ensemble learning, characterized in that: The following steps are involved: Step 1: Extract multiple seismic attributes of the study layer and calculate the reservoir thickness of the well points. After processing and attribute optimization, combine them into sample data, and divide the sample data into an initial training set and an initial test set; Step 2: Input the initial training set into the Stacking ensemble learning base learner for five-fold cross-validation and hyperparameter optimization. The prediction results are combined with the initial training set to obtain the meta-learner training set. At the same time, prediction and performance evaluation are performed on the initial test set. The prediction results are processed to obtain the meta-learner test set. Step 3: Input the meta-learner training set into the stacking ensemble learning meta-learner and perform hyperparameter optimization to obtain a trained meta-learner. Input the multi-attribute data into the trained meta-learner for prediction to obtain the final reservoir thickness prediction result. Use the meta-learner test set to evaluate the performance of the trained meta-learner.
2. The reservoir thickness prediction method based on improved Stacking ensemble learning according to claim 1, characterized in that: The step 1 comprises: extracting a plurality of seismic attributes including bandwidth, number of zero crossings, arc length, root mean square amplitude, energy half time, average energy, average amplitude, main frequency, and positive amplitude from the interpreted layer of the study interval; applying natural potential logging curve data based on the well-seismic relationship and combining it with the logging data to calculate the reservoir thickness of the study interval; combining the well point reservoir thickness data and the multi-attribute data corresponding to the well point into initial sample data; performing a Pearson correlation analysis on the initial sample data, selecting seismic attributes with a high correlation with the reservoir thickness, and obtaining new sample data; and dividing the new sample data into an initial training set and an initial test set.
3. The reservoir thickness prediction method based on improved Stacking ensemble learning according to claim 1, characterized in that: The step 2 includes: dividing the initial training set into five equal parts and inputting the five-fold cross-validation into a stacking ensemble learning base learner composed of a random forest model, an extreme gradient boosting algorithm, and a long short time difference network, while using a Bayesian algorithm to optimize hyperparameters; combining the five-fold cross-validation prediction results of each base learner and merging them with the initial training set to form a training set of a meta-learner; inputting the initial test set into the base learners trained by the five-fold cross-validation for prediction, taking the average of the obtained prediction results as the prediction result of the initial test set of the base learner and performing performance evaluation, and combining the initial test set prediction results of multiple base learners to obtain a test set of the meta-learner.
4. The reservoir thickness prediction method based on improved Stacking ensemble learning according to claim 1, characterized in that: The step 3 includes: inputting the meta-learner training set into a multivariate linear regression model as a Stacking ensemble learning meta-learner, and performing hyperparameter optimization to obtain a trained meta-learner; inputting multi-attribute data into the trained Stacking ensemble learning meta-learner for prediction to obtain a final reservoir thickness prediction result; and using the meta-learner test set to perform performance evaluation on the trained meta-learner.