Permeability prediction using PSO-XGBoost based on logging data

Models for permeability prediction generally can be classified into two major types, physical and fitting models. Universally, physical models are wel-comed by geophysicists since the predicted values are calculated on the basis of logging theory, but they show bad generalization on application due to strict requirements on logging data. Fitting models repre-sented by stepwise regression are capable to make quick prediction, but they are difficult to accurately and analytically explain the relationship between permeability and logging curves because of their cal-culation mechanisms, thus also presenting bad gen-eralization. In order to create a new and more pow-erful fitting model, XGBoost, a widely used fitting model at present, is selected and modified by PSO to optimize hyper-parameter tuning. Then the hybrid model PSO-XGBoost is proposed. In this paper, taking the tight sandstone reservoirs of the Chang 4+5 members as a case, the prediction capability of the PSO-XGBoost mo-del are validated by three well-designed experiments. The experiment results show that: ①Compared with physical models, fitting models utilize a fewer parameters to complete prediction, and present better applicability on permeability prediction when modeling data are insufficient, but they have limits on generalization since the prediction is sensitive to the quality of mode-ling data and thereby usually unstable; ②SVR, GBDT, and XGBoost can be improved by PSO, and the formed PSO-SVR, PSO-GBDT and PSO-XGBoost can figure out permeability rapidly. In comparison, PSO-SVR and PSO-GBDT show relatively unstable prediction due to their sensitivities on the quality of learning samples, while PSO-XGBoost displays better performances in predicting efficiency, reliability of predicted results, and prediction stability. Therefore, PSO-SVR is deemed to be unsuitable on permeability prediction, and PSO-XGBoost suitable; ③The prediction capabilities of stepwise regression, PSO-SVM, PSO-GBDT, and PSO-XGBoost can be enhanced when more learning samples are trained. © 2021, Editorial Department OIL GEOPHYSICAL PROSPECTING. All right reserved.