Pore pressure prediction model based on knowledge perception time fusion converter

Abstract

The invention provides a pore pressure prediction model based on a knowledge perception time fusion converter. According to the model, the prediction accuracy of the pore pressure is effectively improved by integrating field professional knowledge and time sequence data. In the data processing stage, firstly, original logging data are preprocessed, including abnormal value detection, missing value filling and data segmentation, so that the quality and integrity of the data are ensured. In the feature extraction stage, time sequence statistical features, rock physical characteristics and empirical formula parameters are extracted from logging data by using a knowledge extraction method, and the features cover physical parameters, mechanical properties and fluid properties of a stratum. Through feature screening, key features which have great influence on pore pressure prediction are reserved. In the model training and evaluation stage, the model is trained by using training set data, and the generalization ability and robustness of the model are verified on an independent test set. The method not only improves the prediction precision of the model, but also obtains an excellent result in practical application, so that the method becomes an effective choice for replacing the traditional technology.

Description

Technical Field

[0001] The present invention relates to the fields of petroleum engineering and geological engineering, and in particular to a model for predicting formation pore pressure. This model is based on a knowledge-aware time fusion transformer, and by combining domain expertise with time series data, it achieves high-precision prediction of formation pore pressure, and is widely used in the fields of oil and gas exploration, drilling engineering, reservoir evaluation, etc. This technology aims to improve the accuracy and reliability of formation pore pressure prediction, and provide support for the safe exploitation and effective development of oil and gas resources. Background Art

[0002] Formation pore pressure is one of the important parameters affecting oil and gas exploration and development. Accurate prediction of formation pore pressure is crucial to the safety of drilling operations, the development efficiency of oil and gas reservoirs, and the understanding of underground fluid migration. However, formation pore pressure prediction faces many challenges, such as data sparsity, incomplete logging data, and the complexity of geological characteristics.

[0003] Traditional pore pressure prediction methods mainly rely on empirical formulas and physical models. These methods rely on a large amount of historical data and the experience of geological experts to a certain extent, and show certain limitations when dealing with complex geological conditions and high-dimensional time series data. With the development of machine learning and deep learning technologies, more and more studies have begun to try to apply these advanced technologies to improve the accuracy of pore pressure prediction. However, existing machine learning and deep learning models are still insufficient in capturing the unique temporal dynamics of geology and integrating domain knowledge.

[0004] Based on the above background, the present invention proposes a pore pressure prediction model based on a knowledge-aware time fusion transformer. The model effectively improves the accuracy and robustness of pore pressure prediction by deeply integrating geological expertise and time series data. The model uses advanced deep learning algorithms, combines geological knowledge and pore pressure evolution theory, optimizes the parsing and feature extraction process of logging data, and thus achieves high-precision pore pressure prediction.

[0005] The technologies closest to the present invention are:

[0006] (1) The pore pressure prediction method based on support vector machine (SVM) uses statistical learning theory to establish a prediction model, but it is difficult to handle the time series correlation in the data and has poor processing effect on long time series data.

[0007] (2) Pore pressure prediction using long short-term memory (LSTM) networks. This method can process time series data, but has limitations in feature extraction and interpretability, and has high computational complexity.

[0008] However, existing pore pressure prediction methods, such as those in (1) and (2), do not fully apply big data analysis technology to deeply process the original logging data, and traditional BP neural networks usually require a large amount of data for training, making it difficult to obtain models with strong generalization capabilities. In addition, these methods are insufficient in integrating geological expertise and capturing the temporal dynamics of data. Therefore, the pore pressure prediction model based on deep learning needs to comprehensively consider data processing and feature extraction methods, and select appropriate models for training to improve the accuracy and robustness of prediction. Summary of the invention

[0009] In view of the shortcomings of the prior art, the present invention provides a pore pressure prediction method based on a knowledge-aware time fusion transformer. The method integrates geological expertise with time series data and uses deep learning technology to optimize the performance of the prediction model. Specifically, the method includes the following steps:

[0010] Step (1) Data preprocessing: Integrate and clean the logging data, including removing outliers and filling missing values ​​to improve data quality.

[0011] Step (2) Feature engineering: Combined with geological knowledge, time series statistical characteristics, rock physical properties and empirical formula parameters such as compaction trends and porosity changes are extracted from logging data.

[0012] Step (3) Model construction and training: Use the time fusion transformer (TFT) model to process time series dependencies, build a prediction model based on geological knowledge, and train the model using historical data.

[0013] Step (4) Model optimization and evaluation: Use the Bayesian optimization algorithm to tune the model parameters and evaluate the performance and accuracy of the model on an independent test set.

[0014] The technical solution of the present invention is:

[0015] A pore pressure prediction method based on a knowledge-aware time fusion transformer includes a data preprocessing module, a feature engineering module, a model building and training module, and a model optimization and evaluation module, including the following steps:

[0016] Step (1) In the data preprocessing module, the original logging curves are smoothed and filtered to remove the influence of noise and outliers, taking into account the impact of data volume, data quality and data processing operations. Then, the time series statistical characteristics of all logging curves are calculated, including moving average, difference between adjacent data, sliding window variance and skewness, etc., and the rock physical characteristics and empirical parameter characteristics are calculated according to the actual formation conditions.

[0017] Step (2) In the data preprocessing module, the present invention makes full use of the time series characteristics of the logging data, constructs professional field characteristics related to pressure as model input, and introduces physical constraints into the model to better utilize these characteristics. The specific features are as follows:

[0018] (1) Moving average filtering

[0019]

[0020] (2) Adjacent data difference

[0021] Di=X i -X i-1 (2)

[0022] (3) Sliding Window Variance

[0023]

[0024] (4) Sliding Window Skewness

[0025]

[0026] (5) Mud content

[0027]

[0028] (6) Porosity

[0029]

[0030] (7) Eaton experience parameters

[0031]

[0032] (8) Overlying pressure gradient

[0033]

[0034] Step (3) In model construction and training, the model is used to process the time series dependencies in the logging data. The model can effectively capture the long-term and short-term dependencies in the time series through its sequence-to-sequence architecture and self-attention mechanism. During the training process, optimization algorithms such as gradient descent are used to gradually adjust the model parameters so that it can accurately predict the pore pressure. The model features are as follows:

[0035] 1) Gating unit: used to skip unnecessary inputs and can adaptively adjust the network depth and complexity according to the characteristics of the dataset.

[0036] 2) Variable selection network: The variable selection network is used to select the variables most relevant to the prediction target and remove irrelevant noise inputs, which can significantly improve the model prediction performance.

[0037] 3) Static covariate encoder: The static covariate encoder is used to encode static feature information into the network.

[0038] 4) Temporal Fusion Decoder: used to learn the temporal relationship from the input variables.

[0039] 5) Interval prediction: The model predicts different percentiles simultaneously at each time step to get a prediction interval.

[0040] Step (4) In model optimization and evaluation, the model parameters are tuned using the Bayesian optimization algorithm. Bayesian optimization can efficiently find the optimal parameter combination by constructing a probabilistic model of the objective function, thereby improving model performance. The performance and accuracy of the model are evaluated on an independent test set, using evaluation indicators such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R 2 ) etc., to verify the predictive ability and generalization performance of the model.

[0041] Beneficial effects of the present invention:

[0042] (1) The present invention significantly improves the accuracy of pore pressure prediction by integrating geological expertise with time series information in logging data. Traditional prediction methods usually rely only on empirical formulas or simple statistical models and cannot fully utilize the temporal dynamic characteristics of the data. The present invention uses the sequence-to-sequence architecture and self-attention mechanism of the time fusion transformer to effectively capture the long-term and short-term dependencies in the time series, thereby more accurately reflecting the changes in formation pore pressure. This high-precision prediction helps to improve the efficiency of oil and gas exploration and development and reduce drilling risks.

[0043] (2) The variable selection network and static covariate encoder in the present invention enable the model to not only provide high-precision prediction results, but also to reasonably explain these results. The variable selection network significantly improves the prediction performance of the model by selecting the variables most relevant to the prediction target and removing irrelevant noise inputs, while retaining the ability to interpret the prediction results. The static covariate encoder integrates geological background information into the model, allowing the model to take into account the influence of the geological environment when making predictions. This interpretability allows geological experts to better understand the prediction results of the model and enhances the credibility of the model in practical applications.

[0044] (3) The present invention is able to predict different percentiles at each time step, thereby obtaining a prediction interval. This interval prediction capability not only provides a single-point prediction value, but also quantifies the uncertainty of the prediction results. For example, when predicting pore pressure, the model can simultaneously predict pressure values ​​at 50%, 75%, and 90% confidence levels, helping users understand the possible range of prediction results. This is particularly important for decision making because it provides more comprehensive information, which helps to assess risks and develop more cautious and scientific engineering plans. This capability significantly improves the credibility and practicality of prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 It is a structural diagram of the pore pressure prediction method based on the knowledge-aware time fusion transformer of the present invention;

[0047] Figure 2 This is a flow chart of the pore pressure prediction method based on the knowledge-aware time fusion transformer of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] A pore pressure prediction method based on a knowledge-aware time fusion transformer includes a data preprocessing module, a feature extraction and encoding module, and a model training and optimization module, and includes the following steps:

[0050] Step (1) Data collection: Obtain raw data from different logging data sources, including acoustic wave time difference, density, gamma ray and other logging curve data. These data reflect the physical and mechanical properties of the formation and are the basis for pore pressure prediction.

[0051] Step (2) Data cleaning: Perform preliminary processing on the collected data to remove obvious outliers and noise data. Outliers can be identified using statistical methods or machine learning algorithms, such as the Z-score method or the isolation forest algorithm. The purpose of data cleaning is to ensure that the data used in subsequent analysis is reliable and noise-free.

[0052] Step (3) Data balance: To address the problem of data imbalance, oversampling technology is used to adjust data distribution. For example, when the amount of logging data for certain formations is insufficient, SMOTE (synthetic minority oversampling technology) can be used to generate more samples to ensure that the prediction accuracy is not affected by data skew during model training.

[0053] Step (4) Knowledge extraction: Extract time series statistical features, rock physical properties and empirical formula parameters from the cleaned logging data. These features include but are not limited to the compaction trend of the formation, porosity changes, mud content and compression wave velocity. By extracting these features, the physical and mechanical properties of the formation can be more comprehensively reflected.

[0054] Step (5) Feature selection: Evaluate the extracted features and select the key features that have the greatest impact on pore pressure prediction. The purpose of feature selection is to reduce data dimensions and remove redundant information, thereby improving the training efficiency and prediction performance of the model. Commonly used feature selection methods include correlation-based selection and principal component analysis (PCA).

[0055] Step (6) Feature standardization: Standardize all selected features and convert the data to a uniform dimension. The purpose of this step is to eliminate the magnitude differences between different features so that the model can converge better during training. Standardization can use methods such as mean-variance standardization or minimum-maximum standardization.

[0056] Step (7) Model training: The time fusion transformer model is used to process the time series dependencies in the logging data. The model can effectively capture the long-term and short-term dependencies in the time series through its sequence-to-sequence architecture and self-attention mechanism. During the training process, the model is trained using historical logging data, and the model parameters are gradually adjusted through optimization algorithms such as gradient descent.

[0057] Step (8) Parameter tuning: Tune the model parameters using the Bayesian optimization algorithm. Bayesian optimization can efficiently find the optimal parameter combination by building a probabilistic model of the objective function, thereby improving model performance. During the tuning process, the model needs to be repeatedly trained and evaluated to find the best hyperparameter settings.

[0058] Step (9) Model evaluation: Evaluate the performance and accuracy of the model on an independent test set. Use evaluation metrics such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R 2 ) to verify the model's predictive ability and generalization performance. The purpose of the evaluation is to ensure that the model has good stability and reliability in practical applications.

[0059] This paper proposes a pore pressure prediction model based on a knowledge-aware time fusion transformer, which achieves high-precision pore pressure prediction by integrating geological expertise with time series information in logging data. The model demonstrates significant technical advantages in three modules: data preprocessing, feature extraction and encoding, and model training and optimization.

[0060] The data preprocessing module aims to improve data quality and ensure data integrity. First, data cleaning is performed to ensure data accuracy and consistency by removing outliers and filling missing values. Second, in response to the problem of data imbalance, oversampling technology is used to adjust data distribution to improve the effect of model training. These steps ensure that the data used in subsequent analysis and model training is high-quality and balanced.

[0061] The feature extraction and encoding module combines geological knowledge to extract key features from logging data, including time series statistical features, rock physical properties, and empirical formula parameters. Specific features include compaction trends, porosity changes, mud content, and compression wave velocity. These features can fully reflect the physical and mechanical properties of the formation. The feature selection step evaluates the importance of each feature to the prediction model, retains the key features that have the greatest impact on pore pressure prediction, and removes redundant information. Finally, through feature standardization, the data is converted to a unified dimension to optimize the convergence and effect of model training.

[0062] The model training and optimization module is the core of achieving high-precision prediction. First, the time fusion transformer model is used to process the time series dependencies in the logging data. The model can effectively capture the long-term and short-term dependencies in the time series through its sequence-to-sequence architecture and self-attention mechanism. During the model training process, historical logging data is used for training, and the model parameters are gradually adjusted through optimization algorithms such as gradient descent. Subsequently, the model parameters are tuned through the Bayesian optimization algorithm to find the optimal parameter combination and further improve the prediction performance of the model. Finally, the performance and accuracy of the model are verified on an independent test set to ensure the stability and reliability of the prediction results.

[0063] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A pore pressure prediction model based on a knowledge-aware time fusion transformer, comprising a data preprocessing module, a feature extraction and encoding module, and a model training and optimization module, specifically comprising the following steps: Step (1): Considering the completeness, accuracy and consistency of the data, data cleaning is performed, including deleting missing data records and applying outlier detection algorithms to identify and delete abnormal data points. Step (2): In view of the data characteristics of pore pressure prediction, such as data imbalance, oversampling technology is used to adjust the data distribution to enhance the training effect of the model. Step (3): Extract time series statistical features, rock physical properties and empirical formula parameters from the logging data. These features cover the physical parameters, mechanical properties and fluid properties of the formation. Step (4): Use feature selection techniques to evaluate the importance of each feature to the prediction model, remove features with low importance, and retain key features. Step (5): Standardize all features and convert the data to a unified dimension to optimize data processing efficiency in subsequent model training. Step (6): Use the training set data to train the knowledge-aware time fusion transformer model. Step (7): Parameters of the knowledge-aware time fusion transformer model are tuned by implementing a parameter search method based on Bayesian optimization to ensure that the model achieves the optimal prediction effect. Step (8): Verify the generalization ability and robustness of the model on an independent test set.

Images