DATA-DRIVEN MODELS FOR BOTTOMHOLE PRESSURE PREDICTION: APPLICATIONS IN HYDRAULIC FRACTURING, GAS LIFT, AND NITROGEN-LIFTED WELLS

Abstract

Precise forecasting of bottomhole pressure is the heart of optimizing the performance of wells, their operational efficiency, and decision-making in several petroleum engineering applications. Traditional physics-based and empirical correlations are frequently not accurate in complex conditions of multiphase flow, whereas direct downhole measurements are also costly and operationally constrained. This research developed individual machine learning (ML) models for each application to predict bottomhole pressure in hydraulic fracturing, gas lift wells, and nitrogen-lifted wells. Large field data were gathered from vertical wells, consisting of 42 hydraulically fractured wells, 304 gas-lift wells, and 518 nitrogen-lifted wells. A scientific approach was used and included ranking of the features, preprocessing of data, and a series of ML models, such as linear regression, tree-based algorithms, ensemble models, neural networks, and symbolic regression. The standard statistical measures were used to assess model performance and were tested on cross-validation and blind tests on independent wells. Moreover, in the case of nitrogen-lifted wells, the results were compared to five empirical correlations that had been defined. These findings indicate that the accuracy and generalization of the advanced ML models are better than traditional approaches. Symbolic regression also has interpretable equations, which are more transparent. The suggested framework will facilitate cost-efficient, scalable, and real-time monitoring of BHP without downhole gauges, which will facilitate optimization and intelligent field operations.