CLASSIFICATION OF FLOW INSTABILITY IN SUBSEA OIL WELLS: STUDY CASE OF 3W PETROBRAS DATASET

Abstract

This research aims to use machine learning to assist production engineers in classifying flow instability in offshore wells case study. The data used is sourced from Petrobras offshore oil well, namely 3W dataset. The research process begins with conducting Exploratory Data Analysis (EDA) on features relevant to production techniques, followed by calculating classification results using Gaussian Naïve Bayes based on TPOT recommendations and Random Forest Classifier as comparison model. Initially, the data is selected with a focus on normal events and flow instability. Then, the data is cleaned from empty values, divided into training and blind data based on time, and random under-sampling is performed to address class imbalances. Outliers are also removed to enhance data quality. Subsequently, the data is input into the TPOT pipeline for additional feature selection, model selection, and parameter optimization. Research findings indicate that four features used in the model, namely P-TPT, P-MON-CKP, T-JUS-CKP, and P-JUS-CKGL, with the target column Event. Gaussian Naïve Bayes successfully achieves 100% precision and recall in the classification report, while the Random Forest Classifier only achieves 11% precision for flow instability and 51% recall for normal events. Thus, this research concludes that Gaussian Naïve Bayes from TPOT pipeline recommendations outperforms Random Forest Classifier in classifying flow instability.