Análise exploratória preditiva de anomalias em poços de extração de petróleo e gás

Abstract

This work consists of an extensive exploratory data analysis (EDA) and the development of a machine learning model through Deep Learning aimed at optimizing and reducing costs in deepwater oil and gas extraction. The predictive model’s function is to anticipate events that could be costly for the exploration company. To achieve this, data science methods were applied to the 3W dataset created by Petrobras, which contains both real and simulated data from the company’s extraction fields. During the research, transformer and LSTM models were tested, with a comparison between the two, resulting in the selection of the transformer model due to its parallelization capabilities during training. The model evaluation metrics were satisfactory, achieving over 98% accuracy in the final model. Furthermore, interpretability of the model was successfully conducted using the Class Activation Map (CAM) technique, confirming the assessments intuitively and drawing parallels with the physical behavior of the sensors.