Reza Azin

In well design, the key to successfully achieving goals and reducing costs is to design well programs based on the prediction of potential problems. The focus of the drilling industry has shifted towards reducing non-productive and invisible lost times. In this research, using machine learning methods and drilling parameters, the classification of drilling problems has been investigated. The drilling data of 17 wells of one of the offshore fields in southwest of Iran were used to build machine learning models. 411 data sets were extracted from various sources such as daily drilling reports, drilling mud reports, and final well reports after removing outlier data. Using feature selection, the features was reduced from 31 to 14. Using MATLAB R2021b software, multi-class classification models including nearest neighbor, and neural network were implemented. After optimization, the best validation accuracy belonged to the nearest neighbor model with an accuracy of 87.3%. Also, the most accurate models in the testing phase were the nearest neighbor with 86.3%, and the neural network with 80%. Models with appropriate accuracy can be used to predict drilling problems, and subsequently avoid them, while drilling new wells in the field under study.