A Multi-Task Learning Model for Predicting the Number of Diesel Generators on Oil Drilling Rigs

This paper explores five machine learning models to predict the optimal number of required online generators in drilling operations. To begin with, we implement individual, general, and multi-task models that can predict the number of online generators for the current drilling state. In addition, we build a general recursive neural network (RNN) model and a multi-task RNN model that can predict the target for thirty minutes in the future. In our multi-task architecture, we consider each drilling state as a task. Training and test datasets have 60 days of historical data collected from multiple rigs in production. We use a dataset from one rig to train and test the performance of the first three models. The RNN models' training phase uses the same dataset with three sliding windows and the testing phase uses three unseen rig datasets for model testing and performance evaluation. Our results show that the RNN models can predict the targets accurately with recall and f1-score greater than 0.95 on average.