Effective prediction of lost circulation from multiple drilling variables: a class imbalance problem for machine and deep learning algorithms

Abstract Multiple machine learning (ML) and deep learning (DL) models are evaluated and their prediction performance compared in classifying five wellbore fluid-loss classes from a 20-well drilling dataset (Azadegan oil field, Iran). That dataset includes 65,376 data records with seventeen drilling variables. The dataset fluid-loss classes are heavily imbalanced (> 95% of data records belong to the less significant loss classes 1 and 2; only 0.05% of the data records belong to the complete-loss class 5). Class imbalance and the lack of high correlations between the drilling variables and fluid-loss classes pose challenges for ML/DL models. Tree-based and data matching ML algorithms outperform DL and regression-based ML algorithms in predicting the fluid-loss classes. Random forest (RF), after training and testing, makes only 35 prediction errors for all data records. Consideration of precision recall and F1-scores and expanded confusion matrices show that the RF model provides the best predictions for fluid-loss classes 1 to 3, but that for class 4 Adaboost (ADA) and class 5 decision tree (DT) outperform RF. This suggests that an ensemble of the fast to execute RF, ADA and DT models may be the best way to practically achieve reliable wellbore fluid-loss predictions. DL models underperform several ML models evaluated and are particularly poor at predicting the least represented classes 4 and 5. The DL models also require much longer execution times than the ML models, making them less attractive for field operations that require prompt information regarding rapid real-time decision responses to pending class-4 and class-5 fluid-loss events.