(Objective) To improve the accuracy and efficiency of lithology identification and formation parameter prediction during drilling, the study used drilling parameter data and implemented SVM, XGBoost, random forest, LightGBM and an ensemble of five algorithms for comparative experiments under baseline, noise interference (10% and 20% noise levels) and data imbalance scenarios. (Results) In lithology identification, the stacking ensemble achieved accuracies of 86.80% and 83.10% under 10% and 20% noise scenarios respectively. The accuracy of the random forest algorithm was 91.15% for the baseline scenario, and for the imbalanced scenarios (10% and 5%), the accuracies were 88.9% and 88.1%, respectively. In formation parameter prediction, random forest achieved mean absolute errors (MAEs) of 3.45, 0.0019, 0.0008 and 0.0004 for seismic velocity, pore pressure, fracture pressure and overburden pressure in the baseline scenario and performed best under noise and imbalanced data conditions. (Conclusions) An adaptive hybrid model was ultimately established: stacking ensemble is used for lithology prediction in noisy environments, while random forest is used for lithology prediction in non-noisy environments and for formation parameter prediction across all environments.
He et al. (Sat,) studied this question.