digilib@itb.ac.id +62 812 2508 8800

ABSTRAK Melani Ere
PUBLIC Suharsiyah

2022 TA PP MELANI ERE 1.pdf
Terbatas Suharsiyah

One of the ways for drilling optimization is to understand the Rate of Penetration (ROP). There have been many proposed models to estimate the ROP. However, many of them tend to be slow processes and show unsatisfying results. Therefore, the need of a more robust model—more accurate and faster computation using machine learning is proposed and aims to be implemented in real-time. The parameters that affect ROP are Weight On Bit (WOB), Revolution Per Minute (RPM) refers to rotation speed, Torque, and Mud Weight will be input data. This study was primarily concerned with ROP prediction in an oil field at section 6 1/8" and offers two ways of prediction, namely ROP prediction by using data learned from another well and the other by using data learned from the same well but in different depths. The machine learning method will be determined before the focus ROP prediction is performed on Well-3. The data Well-1 and Well-2 will be divided into three parts, 70% training data, 15% validation data, and 15% testing data. The models used are Gradient Boosted Trees (GBT), Random Forest (RF), and Artificial Neural Network (ANN). The model with the highest R2 and the lowest Mean Absolute Percentage Error (MAPE) will be chosen. The best model will be used for further prediction of Well-3. The result is Random Forest has the highest R2 and lowest MAPE. The correlation coefficient was over 0.98 for the training and validation data. For testing data, the correlation coefficient is 0.93, and the MAPE for training, validation, and testing data are 6.1%, 4.9%, and 14%, respectively. The ROP prediction of another well is better than the ROP prediction in different depths at the same well by comparing the correlation result. Although ANN is used in many papers, this paper demonstrates that Random Forest is much faster and more accurate. Random Forest was an appealing option for further ROP prediction or other drilling parameters prediction because it required less time and computational resources than ANN.