Modelos preditivos para ROP como suporte à otimização em tempo real de parâmetros operacionais na perfuração de poços de petróleo

Abstract

The Rate of Penetration (ROP) is a parameter of great interest in real-time drilling optimization. A higher ROP reduces drilling time and can lead to significant cost reductions in well construction. Predictive models for ROP are applied to predict ROP based on measured data while drilling, enabling the determination of optimal operational parameters such as RPM, WOB, and fluid flow when combined with optimization techniques. Obtaining more accurate ROP models is a challenging task due to the large number of factors interacting nonlinearly. This study involves the examination of different ROP models, including traditional ones, like Bourgoyne & Young and Specific Mechanical Energy, and their adaptations, compared with machine learning models such as Artificial Neural Networks (ANN) and Random Forests (RF). Public data from 7 wells were structured into a dataset with relevant information for evaluating the performance of different models in estimating ROP, including operational parameters, drill-bit data, lithology, geophysical logging data, pore pressure gradient, and unconfined compressive strength. In comparative analyses, error metrics such as Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are compared among the different models for each of the 7 wells. Statistical significance analysis is performed with the Bourgoyne & Young model to understand more significant effects on ROP. The interpretability of traditional models, along with hyperparameter tuning, is adopted to employ machine learning models with more meaningful inputs and greater predictive capacity. Next, two strategies found in the literature for using predictive models for ROP in real-time optimization are compared: models trained with offset well data or with the well’s own data, simulating a gradual data acquisition process (continuous learning). The results indicate better performance of machine learning models compared to traditional models. The RF model shows overall better performance in comparative analyses, with smaller errors and lower computational cost. The relevance of torque and the inclusion of formation data (Delta-T compressional) in machine learning models is identified. Also, continuous learning strategy can achieve lower errors, although both strategies are capable of generating appropriate predictions.