| The prediction of total organic carbon content of the reservoir can be used to evaluate the reservoir’s oil and gas level and hydrocarbon generation potential,which is of great significance to unconventional oil and gas exploration and development.However,there are some problems in the traditional logging methods for predicting total organic carbon content in reservoirs,such as poor extraction ability of logging data and limited prediction accuracy.To solve above problems,this thesis takes the total organic carbon content of the reservoir as the research object,takes data processing as starting point,and applies machine learning theoretical methods.From the perspectives of improving logging data quality,logging data features optimization,and total organic carbon content prediction accuracy,research on the prediction method of total organic carbon content based on logging data is carried out systematically.Focusing on solving the problems of redundant logging features and limited prediction accuracy,it provides guidance for improving evaluation capability of reservoir’s oil-gas level and hydrocarbon generation potential evaluation capability.First,in order to improve the quality of logging data,depth correction,data missing and outlier processing,and data normalization processing of original logging data were successively carried out in this thesis,so as to provide effective data support for subsequent feature optimization and total organic carbon content prediction.Further,to solve the problem of logging feature redundancy and irrelevance,this thesis carries out feature optimization research based on embedded feature selection,conducts feature ranking according to the contribution degree of features to the model,studies the influence of contribution threshold on prediction accuracy,and then selects the optimal feature subset.On that basis,this thesis studies and proposes a prediction model of total organic carbon content based on embedded feature selection and gradient boosting regression tree.The model uses embedded feature selection method to extract logging data information effectively,and finally uses the advantages of gradient boosting regression tree algorithm to integrate multiple weak learners,which effectively improves the prediction accuracy of total organic carbon content and reduces the training time cost.The efficiency of proposed method is verified by experimental analysis.The results show that the effect of low quality data on the accuracy of the prediction model is effectively improved by correlation function method,mean-filled missing data,and maximum-minimum standardization method in the research of logging data preprocessing.Further,using embedded feature selection method for feature optimization,the number of input features was reduced from 31 to 3,and the prediction index R~2 achieved 0.992,which significantly reduced data dimension while ensuring prediction accuracy.On that basis,by combining embedded feature selection and gradient boosting regression tree algorithm,the training time of total organic carbon content prediction model was reduced to 0.388s,which effectively improved the prediction speed of total organic carbon content.In addition,compared with the traditional prediction model of total organic carbon content based on neural network,the prediction accuracy R~2 of the model proposed in this paper increased from 0.950 to 0.992,and the training time decreased from 18.799s to 0.388s,effectively improves the performance of the original logging data processing speed and prediction accuracy. |
