Pore pressure real-time prediction method based on stacked integrated model

Abstract

The invention discloses a pore pressure real-time prediction method based on a stacked integrated model. The method comprises the following steps: collecting well logging data of a drilled well in a research block and a cable formation test result; using a Bayesian algorithm to invert a global optimal Eaton index; calculating the pore pressure equivalent density by combining a dc index method and an Eaton method, and establishing a data set; according to the data distribution characteristics of the training set and the test set and the characteristic importance sequence, optimization of input parameter combinations is carried out; constructing a machine learning model based on a stacking integration framework; optimizing the base learner and the meta learner, and optimizing the optimized model hyper-parameters by using a Bayesian optimization algorithm; and testing the prediction performance and generalization ability of the optimized model according to the test set, and evaluating a model prediction result. According to the invention, through a stacking integration strategy, advantage integration is carried out on a plurality of models, and the model hyper-parameters are optimized by using the Bayesian algorithm, so that accurate and real-time prediction of the pore pressure profile can be realized.

Description

Technical Field

[0001] The invention relates to a real-time prediction method for pore pressure based on a stacking integrated model, and belongs to the technical field of petroleum exploration and development. Background Art

[0002] Pore ​​pressure refers to the fluid pressure within the pore spaces of a formation and is primarily influenced by the sedimentary environment, diagenesis, and hydrodynamic evolution. As a key parameter characterizing formation properties, accurate prediction of pore pressure is crucial for optimizing drilling fluid density, avoiding wellbore instability, and optimizing hydraulic fracturing design.

[0003] Pore ​​pressure prediction methods are generally categorized into three types: pre-drilling prediction, in-drilling prediction, and post-drilling evaluation. Pre-drilling prediction primarily relies on seismic interval velocities to predict pore pressure, thereby assisting in well planning. However, due to the limited resolution of seismic data, this method cannot capture localized pressure anomalies, and its accuracy is limited by the quality of the seismic data, introducing significant uncertainty into the prediction. Post-drilling evaluation aims to utilize data from previously drilled wells to refine pore pressure estimates and thus assist in pre-drilling predictions for new wells. Among these, methods based on well log interpretation are the most commonly used. While these methods can provide relatively accurate pore pressure predictions, they are inherently limited to completed wells. During drilling, these methods typically rely on LWD data, which is not always available in every well. LWD prediction enables real-time monitoring of pore pressure, supporting field decision-making and reducing operational risks. Common methods include the DC index method, the sigma logarithm method, and the mechanical specific energy method. These methods use surface logging data such as drilling rate, weight on bit, and torque to estimate pore pressure. However, these empirical methods often find it difficult to dynamically adapt to changes in pore pressure, especially in geologically complex or heterogeneous formations. In addition, frequent changes in bottomhole assembly configurations, continuous advancements in drilling technology, and changes in drilling rig equipment have further reduced the robustness and accuracy of these empirical models. With the continuous development and application of machine learning methods, scholars at home and abroad have begun to use machine learning to predict pore pressure. However, a significant limitation of most existing studies is that they rely on well logging data or logging while drilling data, which limits their application scenarios. The few studies that use only surface logging data for pore pressure prediction usually perform point-to-point predictions, which cannot describe the continuous depth dependence of pore pressure and have limited ability to generate complete and continuous pore pressure profiles, thus limiting their application in actual drilling operations.

[0004] Currently, there is no report on a method for real-time prediction of pore pressure profiles based on a stacked integrated model using conventional surface logging data as input parameters. Summary of the Invention

[0005] The purpose of the present invention is to address the problems existing in the prior art and provide a real-time prediction method for pore pressure based on a stacked integrated model. This method integrates the advantages of multiple models through a stacked integration strategy, and uses a Bayesian algorithm to optimize the model hyperparameters, which can achieve accurate and real-time prediction of the pore pressure profile.

[0006] The present invention solves the above technical problems by providing a technical solution: a real-time pore pressure prediction method based on a stacked integrated model, comprising the following steps:

[0007] Step S1: collecting logging data and wireline formation test results of wells drilled in the study area;

[0008] Step S2: Based on the cable formation test results, the global optimal Eaton index is obtained by inversion using the Bayesian algorithm;

[0009] Step S3: Calculate the pore pressure equivalent density based on the global optimal Eaton index and combine the DC index method and the Eaton method to establish a data set;

[0010] Step S4: normalize the input parameters and divide the data set into a training set and a test set;

[0011] Step S5: Optimize the input parameter combination according to the data distribution characteristics and feature importance ranking of the training set and the test set;

[0012] Step S6: building a machine learning model based on a stacking integration framework;

[0013] The machine learning model based on the stacked inheritance framework includes a base learner and a meta learner;

[0014] Step S7: Optimize the base learner and meta-learner based on the prediction effect evaluation index, and then use the Bayesian optimization algorithm to optimize the hyperparameters of the optimized model;

[0015] Step S8: Test the prediction performance and generalization ability of the optimized model based on the test set, and evaluate the model prediction results through evaluation indicators.

[0016] A further technical solution is that the surface logging data includes well depth, vertical depth, mechanical penetration rate, drilling pressure, hook load, rotation speed, torque, standpipe pressure, drilling fluid density, total gas content, drill bit footage, drilling time, drilling fluid inflow flow rate, drilling fluid outflow flow rate, hook height, and pumping time.

[0017] A further technical solution is that the specific process of step 2 is:

[0018] Step S21: randomly set an initial Eaton index m0, and calculate the pore pressure equivalent density according to the DC index method and the Eaton method;

[0019] Step S22, calculating the residual Loss between the pore pressure equivalent density in step S21 and the actual pore pressure value obtained by the wireline formation test;

[0020] Step S23: Use the Bayesian algorithm to iteratively update the Eaton index, thereby continuously reducing the Loss value, and stop the iteration when the set Loss threshold or the maximum number of iterations is reached, thereby obtaining the global optimal Eaton index.

[0021] A further technical solution is that the calculation formula in step S21 is:

[0022]

[0023] Where: d c is the dc index, dimensionless; N is the rotation speed, r / min; R is the mechanical drilling speed, m / min; W is the drilling pressure, kN; D b is the drill bit diameter, mm; ρ n is the normal mud density, g / cm 3 ρ d is the actual drilling fluid density, g / cm 3 ρ p is the pore pressure equivalent density, g / cm 3 ρ o is the pressure equivalent density of the overburden, g / cm 3 ;d cn is the dc index corresponding to the normal compaction trend line, dimensionless; m is the Eaton index, dimensionless.

[0024] A further technical solution is that the calculation formula in step S22 is:

[0025]

[0026] Where: Loss is the residual, g / cm 3 ; is the actual pore pressure equivalent density, g / cm 3 ρ p is the pore pressure equivalent density, g / cm 3 .

[0027] A further technical solution is that the specific process of step S5 is:

[0028] Step S51: based on the data distribution characteristics between the training set and the test set, remove features with large distribution differences;

[0029] Step S52: perform feature importance analysis based on the SHAP value, select features with higher feature importance, and determine the optimal input parameter combination.

[0030] A further technical solution is that the SHAP value calculation formula can be expressed as:

[0031]

[0032] Where: φ i is the feature SHAP value, dimensionless; S is the feature set, dimensionless; M is the number of features, dimensionless; f x is the predicted value of pore pressure equivalent density, g / cm 3 ; i is the i-th eigenvalue, dimensionless.

[0033] A further technical solution is that the base learners include: ridge regression, decision tree, random forest, adaptive gradient boosting tree, extreme gradient boosting tree, lightweight gradient boosting tree, BP neural network, convolutional neural network and long short-term memory neural network; the meta-learners include: ridge regression, decision tree, random forest, adaptive gradient boosting tree, extreme gradient boosting tree.

[0034] A further technical solution is that the evaluation indicators include mean absolute error, root mean square error, absolute percentage error, mean absolute percentage error, and goodness of fit.

[0035] The present invention has the following beneficial effects: the present invention adopts a stacked integrated machine learning model to perform real-time prediction of pore pressure; this method takes into account the data distribution differences between the training set and the test set, and integrates the advantages of multiple models, further improving the prediction accuracy and generalization ability of the model, which plays a very important role in drilling process design, hydraulic fracturing design, and drilling fluid design. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the Eaton index inversion result diagram;

[0037] Figure 2 This is the calculation result diagram of pore pressure equivalent density;

[0038] Figure 3 This is a feature selection result diagram based on sample distribution characteristics;

[0039] Figure 4 This is the result graph of feature selection based on SHAP value;

[0040] Figure 5 Schematic diagram for stacking integrated machine learning models;

[0041] Figure 6Evaluation index diagram for prediction results of different base learners;

[0042] Figure 7 Evaluation index diagram for prediction results of different meta-learners;

[0043] Figure 8 This is a comparison chart of model prediction results before and after hyperparameter optimization;

[0044] Figure 9 This is a map of the spatial distribution of drilled wells in the block;

[0045] Figure 10 This is the prediction result and evaluation index result diagram of the test well in scenario 1;

[0046] Figure 11 This is the test well prediction result and evaluation index result diagram under scenario 2;

[0047] Figure 12 This is the prediction result and evaluation index result diagram of the test well under scenario three. DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0049] The present invention provides a real-time pore pressure prediction method based on a stacked integrated model, comprising the following steps:

[0050] Step S1: collecting logging data and wireline formation test results of wells drilled in the study area;

[0051] In this example, surface logging data of 6 wells in the study area were collected, as shown in Table 1, and the wireline formation test results are shown in Table 2;

[0052] Table 1 Parameter statistical characteristics

[0053]

[0054] Table 2 Pore pressure measurement results

[0055]

[0056]

[0057] Step S2: Based on the cable formation test results, the global optimal Eaton index is obtained by inversion using the Bayesian algorithm;

[0058] Step S21: randomly set an initial Eaton index m0, and calculate the pore pressure equivalent density according to the DC index method and the Eaton method;

[0059] The DC index calculation formula can be expressed as:

[0060]

[0061] Where: d c is the dc index, dimensionless; N is the rotation speed, r / min; R is the mechanical drilling speed, m / min; W is the drilling pressure, kN; D b is the drill bit diameter, mm; ρ n is the normal mud density, g / cm 3 ρ d is the actual drilling fluid density, g / cm 3 ;

[0062] The formula for calculating pore pressure using the Eaton method can be expressed as:

[0063]

[0064] Where: ρ p is the pore pressure equivalent density, g / cm 3 ρ o is the pressure equivalent density of the overburden, g / cm 3 ;d cn is the dc index corresponding to the normal compaction trend line, dimensionless; m is the Eaton index, dimensionless;

[0065] Step S22, calculating the residual Loss between the pore pressure equivalent density in step S21 and the actual pore pressure value obtained by the wireline formation test;

[0066]

[0067] Where: Loss is the residual, g / cm 3 ; is the actual pore pressure equivalent density, g / cm 3 ρ p is the pore pressure equivalent density, g / cm 3 ;

[0068] Step S23: Use the Bayesian algorithm to iteratively update the Eaton index, thereby continuously reducing the Loss value, and stop the iteration when the set Loss threshold or the maximum number of iterations is reached, thereby obtaining the global optimal Eaton index;

[0069] The inversion results are as follows Figure 1 As shown, the global optimal Eaton index of this block is m = 0.2449;

[0070] Step S3: Calculate the pore pressure equivalent density based on the global optimal Eaton index and in combination with the DC index method and the Eaton method, and establish a data set using the logging data as input parameters and the pore pressure equivalent density as output parameters;

[0071] Step S4: normalize the input parameters and divide the data set into a training set and a test set;

[0072]

[0073] Where: X * is the normalized input parameter; X max and X min are the maximum and minimum values ​​of the input parameters respectively;

[0074] Step S5: Optimize the input parameter combination according to the data distribution characteristics and feature importance ranking of the training set and the test set;

[0075] Step S51: based on the data distribution characteristics between the training set and the test set, remove features with large distribution differences;

[0076] The input parameters of the training set and the test set are fed into a random forest classifier. The purpose of this classifier is to distinguish whether the input data belongs to the training set or the test set, and to sort the contribution of the features in the classification results. Furthermore, through ablation experiments, the input features are removed one by one in the order of data distribution difference from large to small, and the input parameter combination with the best prediction effect is retained. The prediction effect of the model under different input parameter combinations is shown in the figure below. Figure 3 As shown;

[0077] Step S52: perform feature importance analysis based on the SHAP value, select features with higher feature importance, and determine the optimal input parameter combination;

[0078] The SHAP value calculation formula can be expressed as:

[0079]

[0080] Where: φ i is the feature SHAP value, dimensionless; S is the feature set, dimensionless; M is the number of features, dimensionless; f x is the predicted value of pore pressure equivalent density, g / cm 3 ; i is the i-th eigenvalue, dimensionless.

[0081] The results of parameter selection based on feature importance ranking are as follows Figure 4As shown, the final input parameter combination is: R-ROP, WOB, TG, PT, MW, PFI, TVD, H, BR, BT, and SPP.

[0082] Step S6: Construct a machine learning model based on a stacking integration framework (the principle is as follows Figure 5 shown);

[0083] The machine learning model based on the stacked inheritance framework includes a base learner and a meta learner;

[0084] The base learners include: Ridge Regression (Ridge), Decision Tree (DT), Random Forest (RF), Adaptive Gradient Boosting Tree (AdaBoost), Extreme Gradient Boosting Tree (XGBoost), Lightweight Gradient Boosting Tree (LightGBM), BP Neural Network (BPNN), Convolutional Neural Network (CNN) and Long Short-Term Memory Neural Network (LSTM); the meta-learners include: Ridge Regression (Ridge), Decision Tree (DT), Random Forest (RF), Adaptive Gradient Boosting Tree (AdaBoost), Extreme Gradient Boosting Tree (XGBoost);

[0085] Step S7: Optimize the base learner and meta-learner based on the prediction effect evaluation index, and then use the Bayesian optimization algorithm to optimize the hyperparameters of the optimized model;

[0086] The calculation formula of the evaluation index used to evaluate the prediction effect of the model can be expressed as:

[0087]

[0088] Where: MAE is the mean absolute error, g / cm 3 ; RMSE is root mean square error, g / cm 3 ; APE is absolute percentage error, %; MAPE is mean absolute percentage error, %; R 2 is the goodness of fit, dimensionless.

[0089] The evaluation indicators of the prediction results of different base learners are as follows: Figure 6 As shown in the figure, the base learners finally selected are CNN, LSTM, BPNN, and LightGBM.

[0090] The meta-learner is optimized based on the prediction effect, and the prediction result evaluation indicators of different meta-learners are as follows: Figure 7 As shown, the meta-learner finally selected is XGBoost;

[0091] The hyperparameter ranges of the base learner and meta-learner are set, and the Bayesian optimization algorithm is used to optimize the hyperparameters of each learner. The hyperparameter ranges of the learners and the final selected hyperparameters are shown in Table 3.

[0092] Table 3 Hyperparameter optimization results

[0093]

[0094]

[0095] The pore pressure is predicted on the validation set using the model after hyperparameter optimization. The comparison of the prediction results of the model before and after optimization is shown in the figure below. Figure 8 As shown;

[0096] Step S8: Test the prediction performance and generalization ability of the optimized model based on the test set, and evaluate the model prediction results through evaluation indicators.

[0097] Step S81: The spatial distribution of the wells drilled in the block is as follows: Figure 9 As shown in the figure, three different drilling scenarios are set according to the spatial relationship between the assumed drilling well and the adjacent wells that have been drilled.

[0098] Scenario 1: There are completed adjacent wells in all directions around the drilling well. In this scenario, the test well is well X1, and the training wells are wells X2, X3, X4, X5, and X6.

[0099] Scenario 2: The drilling well has completed adjacent wells in only one direction. In this scenario, the test well is Well X5, and the training wells are Wells X1, X2, X3, X4, and X6.

[0100] Scenario 3: There are no completed wells nearby the drilling well. In this scenario, the test well is Well X3, and the training wells are Wells X1, X2, X4, X5, and X6.

[0101] Step S82: Substitute the data in the training set under different scenarios into the optimized stacked ensemble model to train the models separately;

[0102] Step S83: Substitute the data in the test set under different scenarios into the trained model to predict and obtain pore pressure prediction data;

[0103] Step S84: Calculate an evaluation index based on the pore pressure prediction data and the actual pore pressure data in the test set, and then evaluate the prediction effect.

[0104] The comparison diagram of the pore pressure profile of the model prediction results obtained in steps S81 to S84 and the actual value of the pore pressure, as well as the evaluation indicators such as Figure 10-12As shown in the figure, the prediction results show that when there are many completed wells around the target well, the data distribution difference between the training set and the test set is small, so the model can achieve good prediction accuracy. As the number of neighboring wells near the target well decreases, the data distribution difference between the training set and the test set gradually increases, and the model prediction performance gradually decreases. In terms of prediction accuracy, the model prediction accuracy is the highest in scenario 1, followed by scenario 2, and the model prediction accuracy for scenario 3 is relatively low. The comprehensive results show that the method proposed in this study can achieve accurate and real-time pore pressure profile prediction.

[0105] The above description does not limit the present invention in any form. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any technician familiar with the profession can use the technical content disclosed above to make some changes or modifications to equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are within the scope of the technical solution of the present invention.

Claims

1. A real-time prediction method for pore pressure based on a stacked integrated model, characterized in that: The following steps are involved: Step S1: collecting logging data and wireline formation test results of wells drilled in the study area; Step S2: Based on the cable formation test results, the global optimal Eaton index is obtained by inversion using the Bayesian algorithm; Step S3: Calculate the pore pressure equivalent density based on the global optimal Eaton index and combine the DC index method and the Eaton method to establish a data set; Step S4: normalize the input parameters and divide the data set into a training set and a test set; Step S5: Optimize the input parameter combination according to the data distribution characteristics and feature importance ranking of the training set and the test set; Step S6: building a machine learning model based on a stacking integration framework; The machine learning model based on the stacking inheritance framework includes a base learner and a meta learner; Step S7: Optimize the base learner and meta-learner based on the prediction effect evaluation index, and then use the Bayesian optimization algorithm to optimize the hyperparameters of the optimized model; Step S8: Test the prediction performance and generalization ability of the optimized model based on the test set, and evaluate the model prediction results through evaluation indicators.

2. The real-time pore pressure prediction method based on the stacked integrated model according to claim 1 is characterized in that: The surface logging data includes well depth, vertical depth, mechanical penetration rate, drilling pressure, hook load, rotation speed, torque, standpipe pressure, drilling fluid density, total gas content, drill bit footage, drilling time, drilling fluid inflow rate, drilling fluid outflow rate, hook height, and pumping time.

3. The real-time prediction method for pore pressure based on the stacked integrated model according to claim 1 is characterized in that: The specific process of step 2 is: Step S21: randomly set an initial Eaton index m0, and calculate the pore pressure equivalent density according to the DC index method and the Eaton method; Step S22, calculating the residual Loss between the pore pressure equivalent density in step S21 and the actual pore pressure value obtained by the wireline formation test; Step S23: Use the Bayesian algorithm to iteratively update the Eaton index, thereby continuously reducing the Loss value, and stop the iteration when the set Loss threshold or the maximum number of iterations is reached, thereby obtaining the global optimal Eaton index.

4. The real-time pore pressure prediction method based on the stacked integrated model according to claim 3 is characterized in that: The calculation formula in step S21 is: Where: d c is the dc index, dimensionless; N is the rotation speed, r / min; R is the mechanical drilling speed, m / min; W is the drilling pressure, kN; D b is the drill diameter, mm; ρ n is the normal mud density, g / cm 3 ρ d is the actual drilling fluid density, g / cm 3 ρ p is the pore pressure equivalent density, g / cm 3 ; ρ o is the pressure equivalent density of the overlying rock, g / cm 3 ;d cn is the dc index corresponding to the normal compaction trend line, dimensionless; m is the Eaton index, dimensionless.

5. The real-time prediction method for pore pressure based on the stacked integrated model according to claim 3 is characterized in that: The calculation formula in step S22 is: Where: Loss is the residual, g / cm 3 ; is the actual pore pressure equivalent density, g / cm 3 ρ p is the pore pressure equivalent density, g / cm 3 .

6. The real-time prediction method for pore pressure based on the stacked integrated model according to claim 1, characterized in that: The specific process of step S5 is as follows: Step S51: based on the data distribution characteristics between the training set and the test set, remove features with large distribution differences; Step S52: perform feature importance analysis based on the SHAP value, select features with higher feature importance, and determine the optimal input parameter combination.

7. The real-time prediction method for pore pressure based on the stacked integrated model according to claim 6, characterized in that: The SHAP value calculation formula can be expressed as: Where: φ i is the feature SHAP value, dimensionless; S is the feature set, dimensionless; M is the number of features, dimensionless; f x is the predicted value of pore pressure equivalent density, g / cm 3 ; i is the i-th eigenvalue, dimensionless.

8. The real-time prediction method for pore pressure based on a stacked integrated model according to claim 1, characterized in that: The base learners include: ridge regression, decision tree, random forest, adaptive gradient boosting tree, extreme gradient boosting tree, lightweight gradient boosting tree, BP neural network, convolutional neural network and long short-term memory neural network; the meta learners include: ridge regression, decision tree, random forest, adaptive gradient boosting tree, extreme gradient boosting tree.

9. The real-time pore pressure prediction method based on the stacked integrated model according to claim 1 is characterized in that: The evaluation indicators include mean absolute error, root mean square error, absolute percentage error, mean absolute percentage error, and goodness of fit.

Images