Logging Parameter Optimal Selection For Reservoir Permeability Intelligent Prediction

The theoretical research and technical development of reservoir permeability prediction based on conventional logging data are of great significance to reservoir evaluation and oil-gas development.The existing methods have certain limitations: 1)Input features(logging parameters)are selected based on expert experience,some of which cannot be obtained directly from logging.2)The prediction accuracy and efficiency of conventional permeability prediction model based on machine learning algorithm are low.3)There is no effective solution to the cross-well prediction problem caused by the difference of data distribution.To solve the above problems,this thesis takes an oil field in western China as the research object by using machine learning technology,and carries out research on reservoir permeability intelligent prediction systematically through logging data feature selection,ensemble learning permeability prediction and transfer learning cross-well prediction.It aims to solve the problems of conventional methods relying on expert experience,low prediction accuracy and cross-well permeability prediction.This improves the accuracy and efficiency of permeability prediction to a certain extent.To reduce the dependence of input feature selection on expert experience,the original logging parameter selection methods are studied.The performance of filtered,wrapper and embedded feature selection method is analyzed and compared.It is decided to use the embedded feature selection method to provide the optimal feature subset for the subsequent reservoir permeability prediction.Further,a reservoir permeability intelligent prediction model based on Light GBM is designed and constructed using the ensemble learning ideas to improve the efficiency and accuracy of reservoir permeability prediction.It improves the permeability prediction accuracy.On this basis,a sample transfer learning method based on EDFSW is studied and proposed to solve the problem of data drift and distribution difference in cross-well permeability prediction.It can use the source domain samples to predict the target domain permeability and improve the accuracy of cross-well permeability prediction.The efficiency of the proposed method is verified by experimental analysis.In the analysis and comparative study of logging parameter selection,the embedded feature selection method effectively reduces the input feature dimension without relying on expert experience,and can achieve satisfactory prediction effect by using the least logging parameters.Compared with other permeability prediction models,The permeability prediction model based on Light GBM can effectively improve the permeability prediction efficiency and accuracy because of its fast training speed and good prediction effect.The proposed sample transfer learning algorithm based on EDFSW can effectively filter out source domain samples similar to the target domain to improving the cross-well permeability prediction effect.