Predicting Rig Engine Utilization Using Machine Learning to Drive Fleet-Wide Emissions Reductions Initiatives

Abstract

Abstract With the drilling industry's focus on sustainability and digitalization, there are several data-driven studies published in the recent years which aim to reduce the operational fuel consumption and greenhouse gas (GHG) emission footprint through use of efficient generator management and alternative fuel sources. The possible savings from these optimization solutions are often measured with respect to a baseline - a model representing expected performance for the observed load profile. Since rig power demand and crew behaviors can vary immensely, estimating this baseline accurately can be challenging, while still being critical for GHG reduction initiatives to measure progress. This paper presents a novel data-driven methodology using machine learning to provide a real-time representative model of generator management. Two years of generator power output data at 1 Hz sampling frequency was collected over 110 rigs for this study. From this data, multiple time-dependent features were created to capture and learn from trends in rig power requirements. An ensemble of rig-specific models was built using an ordinal classification algorithm to predict the number of generators online at any particular instant, while remaining true to the core behaviors across the fleet. Additional logic was added during testing to ensure results were realistic and interpretable by field personnel. This real-time model was then deployed to run on this contractor's edge-computing platform across its fleet of rigs. Emissions metrics using the proposed methodology were one of the crucial results that encouraged the rig crews to reduce the fuel consumed over the duration of this study. These metrics were integrated into proactive reporting sent to rigs to improve their generator management practices and to highlight rigs with outstanding practices. Results on the impact of these initiatives on performance will be discussed. The authors are not currently aware of any large-scale, data-driven approaches to estimating generator usage on drilling rigs. This approach offers new ways to assist and make an impact on fleet-wide GHG reduction initiatives. Additionally, the unique ensemble framework of this methodology can be customized and adapted easily to many applications involving timeseries prediction over multiple rigs / operating units. This paper also demonstrates the value of data monitoring and data-driven workflows in increasing the operational efficiency.