Research On The Method Of Predicting Shale Reservoir Parameters In The Qingshankou Formation Of Well B Area In The Songliao Basin

Reservoir parameters are crucial for calculations of oil and gas reserves,and evaluations of reservoir properties.Their accuracy and reliability are of utmost importance.Due to the limitation in the number of core samples,core analysis usually yields relatively limited data on reservoir parameters,while a wealth of logging data can be obtained in the process of oil and gas exploration and development.How to scientifically use logging data to improve the accuracy of reservoir parameter prediction is one of the current research hotspots.With the development of computer technology,machine learning and deep learning methods have become increasingly mature and have achieved significant results in various fields.This thesis is based on machine learning and deep learning methods to study the prediction method of shale reservoir parameters in the Qingshankou Formation,B Well Area,Songliao Basin.The specific research content is as follows:(1)For problems such as outliers or missing values in the data,imbalanced data distribution,and empirical risk in feature selection during model construction,this thesis first uses data preprocessing techniques to remove outliers in the data and fill in missing values.Then,data visualization technology is used to analyze the data,and features with data bias are processed.Finally,feature selection is performed using the correlation coefficient method and tree model feature contribution method in statistics.(2)To address the issue of monotonous reservoir parameter prediction methods under conventional sample volume conditions,this thesis uses the Stacking ensemble learning method for modeling.The base learners include neural networks,Bagging,and Boosting representative algorithms,namely,BP neural networks,random forests,and XGBoost,three different types of algorithms.The meta-learner uses ridge regression algorithms.The results show that the RMSE of BP neural network,random forest,and XGBoost on the test set are 0.44,0.24,and0.22,respectively.The RMSE of the Stacking ensemble learning method is 0.20,which is better than other models.At the same time,the accuracy of this method has increased by approximately 26% compared to Δlg R.To prove the reliability of the method,the permeability is modeled according to this process.The result is close to the experimental result,with small error and high stability.Therefore,the thesis believes that using the Stacking ensemble learning method for reservoir parameter prediction under conventional sample volume conditions has higher accuracy and reliability.(3)For some reservoir parameters with small data volume,it is difficult to obtain highprecision models with conventional machine learning methods.This thesis proposes a GBDTTab Net model,which introduces the Tab Net model into the field of reservoir parameter prediction for the first time,expands the dataset using the concept of semi-supervised learning,and establishes a high-precision model using the GBDT algorithm.This model generates a large number of pseudo-labels and finally uses the Tab Net network for modeling.To improve the stability of the model,this thesis has improved the Tab Net network structure by adding Dropout modules and residual networks to suppress model overfitting.The results show that the improved model has a significant improvement in prediction accuracy,with RMSE on the test set decreased by 9%,and the accuracy has relatively improved by about 20% compared to the traditional polynomial regression method.To prove the reliability of the method,saturation is modeled according to this process.The result is close to the experimental result,with small error and high stability,which proves the effectiveness of this method.(4)Based on the above research content and results,an intelligent reservoir parameter prediction system with data preprocessing,feature selection,model construction,parameter optimization,and reservoir parameter prediction functions has been designed and developed using open-source libraries such as Py Qt5,Sklearn,and Py Torch.This system realizes efficient and high-accuracy prediction of shale reservoir parameters.In conclusion,this thesis posits that the aforementioned research findings offer novel approaches and methodologies for predicting reservoir parameters of the Qingshankou Formation in the B Well Area of the Songliao Basin,and bear significant implications for propelling oil and gas exploration and development in the study area.