Lithology prediction of tight sandstone formation using GS-LightGBM hybrid machine learning model

Abstract

Classic lithology predictors, represented by crossplot, are generally ineffective for tight sandstone formation, mainly due to a point that most lithologies present extremely similar logging responses and thus are rather difficult to be analyzed effectively via crossplot.Compared to classic pattern recognizers, LightGBM shows higher efficiency in data process, therefore it is employed to make a solution for lithology prediction of tight sandstone formation.As LightGBM utilizes many hyper-parameters during modeling, easily causing an issue that the predicted results are not reliable enough, GS algorithm is adopted to solve optimization and further a hybrid machine learning model GS-LightGBM is proposed.The tight sandstone formation of member of Chang 4+5 in western Jiyuan Oilfield is validation targets, and two experiments are designed to reveal prediction capability of the proposed model.In order to highlight validation effect, SVM and XGBoost are introduced as comparative predictors.Experimental results manifest GS-XGBoost and GS-LightGBM have the similar and also the highest marks in the prediction performance of accuracy, F1-score, and AUC, while computing time of GS-LightGBM is only 1/23 shorter than that of GS-XGBoost.The results demonstrate the proposed model is capable to rapidly figure out the predicted lithologies based on guarantee of high prediction accuracy, proving its better applicable prospect and generalization in the study field of lithology prediction of tight sandstone formation.

Keywords

• tight sandstone formation
• lithology prediction
• svm model
• xgboost model
• lightgbm model
• gs optimizing algorithm