Predicción de problemas en el proceso de perforación de pozos petroleros aplicando aprendizaje de máquina supervisado : validación del modelo de aprendizaje de máquina a partir una base de datos.

Abstract

Machine learning algorithms are employed to analyze and process extensive sets of information with the goal of making data-driven decisions, achieving accurate predictions, and automating complex tasks across a range of fields. The petroleum industry harnesses this data-driven learning capability to optimize processes, troubleshoot issues, and efficiently gain valuable insights. This research verified the accuracy of two models developed using XGBoost based on the information extracted from daily drilling reports, final drilling fluid reports, and lithological records (masterlog) from 104 wells in Block 60 Sacha Field. The test dataset comprises 51 cases spanning normal conditions, stuck pipe incidents, influxes, lost circulation, mechanical pack-offs, and blockages, aiming to assess the performance of XG Boost 1 and XG Boost 2 models. In the former case, the model achieved a prediction rate of 88%, with 6 misclassified instances. Conversely, the latter model achieved an 80% accuracy rate, with 10 misclassifications. The adjustment of hyperparameters significantly influences learning accuracy; however, other factors such as computational speed and capacity play critical roles in model selection. Hence, it was concluded that XG Boost 1 is better suited for real data prediction. This preference is based not only on its higher prediction accuracy compared to the second model but also on its use of a higher learning rate, which reduces computational resource demands.