This paper will review the different types of artificial lift problems that have been diagnosed and predicted using machine learning algorithms including tubing and pump failure along with suboptimal performance of sucker rod pumps and and gas injection. The discussion will review the suboptimal performance periods that can be diagnosed and the associated impact on well performance. The presentation will focus on artificial lift and specifically on sucker rod pumps and gas lift.
The machine learning models were trained using production data for over 1 year from operating wells in the Bakken. Time series sensor and controller data for 800 wells was loaded into an AWS cloud platform for analysis and profiling. Experts in each lift type reviewed failure and performance data and identified periods of failure modes and sub-optimal performance. These events were used to train different types of machine learning and artificial intelligence algorithms include random forest, gradient boosted tree and neural network models. These models were used to diagnose and predict the problem states. The overall performance of the models was analyzed to rank and assess performance of each model, lift type and problem state.
Data science, machine learning and artificial intelligence are revolutionizing industries by applying massive low-cost computing power to optimize machine performance. Artificial lift is a good use case given the increasing number of sensors and signals being produced by the latest controllers, the improving data connectivity, lower cost of compute and data storage and the complexity of the surface and downhole problems in unconventional wells. The tubing and pump failure models proved to be highly accurate with diagnostic model accuracy > 99% and precision ranging from 50% to 60%. The long-term tubing failure model performance achieved an expected estimation error of just 31 days on a range of asset lifespans from 100 to 600 days. Suboptimal performance models had accuracy > 90%. The model performance enables high confidence interaction with the wells to react more quickly to problem states and thereby reduce downtime and improve pump effectiveness.
One of the key challenges in deploying data science in the field is to facilitate cooperation between engineers, operators and data scientists. Most of the advanced models used in other industries are based on data from on-line systems (e.g. Google, Facebook) with consistent data and defined outcomes. The oil field has unique data and environmental challenges. This paper discusses how experts in the artificial lift interacted with data scientists to understand this dynamic operating environment, the challenges with data quality and how these assets operate in the field.