Drilling optimization of petroleum and natural gas wells: application of artificial intelligence

• Ano de defesa: 2019
• Autor(a) principal: Barbosa, Luís Felipe Ferreira Motta
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Estadual Paulista (Unesp)
• Palavras-chave em Português: Multi-objective optimization; Drilling efficiency; Rate of penetration; Machine learning; Random forest; Otimização multiobjetiva; Eficiência da perfuração; Taxa de penetração; Aprendizagem de máquinas; Florestas aleatórias; Máquinas de perfuração; Poços de petróleo - Perfuração; Gás natural; Indústria petrolífera
• Link de acesso: http://hdl.handle.net/11449/180784

To meet the increasing primary energy demand, more challenging petroleum reserves started being explored, such as the reservoirs from pre-salt formation close to the Brazilian and Angolan coasts. Historically, low penetration rates in drilling the pre-salt carbonates were reported in the literature, resulting in large capital expenditure on well’s construction. Since the major part of exploration cost is associated with drilling, optimizing this activity is of major importance. In this context, the main objective of the present thesis is to investigate methods for real-time drilling optimization of oil and natural gas wells. A common way to optimize drilling activities is to determine the optimum operational variables (e.g. weight-on-bit and rotational speed) that maximizes the ROP. However, this may yield a decrease in drilling efficiency. An alternative to reduce problems related to drilling inefficiency, such as excessive bit wear and vibrations, is through the selection of operational variables able to minimize the specific energy (SE) spent to excavate a volumetric unit of rock. For that, it is necessary to employ accurate predictive models able to capture how the operational variables (weight-on-bit, rotational speed, mud flow and so on) influence not only on ROP but also on SE. Therefore, the present thesis employed a well-known machine learning method, called random forest, instead of analytical equations found in drilling engineering books. Thus, it was possible to obtain accurate predictive models for ROP and SE, to be used, later, as objective functions in optimization problems to determine the optimum parameters, weight-on-bit and rotational speed. Real-time drilling data from pre-salt region and Norwegian continental shelf were employed. First, several aspects related to training process of random forests were investigated. Among them, it was confirmed the possibility of predicting the ROP with accuracy by employing only four inputs: depth, weight-on-bit, rotational speed, and mud flow. The prediction of SE was carried out by coupling the mathematical formulation with predictive models of ROP and torque (if available). Optimization problems were analyzed with one objective function, as well as with multiple objective functions through the ε-constraint technique. It was observed the sole maximization of ROP may lead to increase in the energy required to drill. However, by imposing the inequality SE(x) ≤ SEacutal*ε on the maximization of ROP, it was possible to reduce significantly the amount of observations whose ROP increased due to detriment of drilling efficiency. For the minimization of SE problems, it was observed a special care to be taken when simulating low-values for weigh-on-bit and rotational speed.