Improved Operational Carbon Footprint Reduction by Using a Data-Driven Method in Chemical Treatment Candidate Selection

Abstract

Abstract This paper presents a physics based and data driven machine learning approach for chemical treatment candidate well selection in fields producing heavy oil. In heavy oil fields, cyclic steam is often used to not only to stimulate the formation around a wellbore but also to open the casing perforations or liner slots plugged by asphaltene deposits. These asphaltene and paraffin build ups can partially or fully block the passages to fluid production. Such cyclic steam jobs are done when there is significant decline in production as measured with usually a threshold over 60 days of operation. But generating steam and using it in cyclic steam jobs is costly and contribute to the carbon footprint. An alternative solution is to employ chemical treatment of wells to clean up blockages that may occur in formations (Brownlee and Zern, 1997). When there is true blockage due to deposits by heavier oil components, cleaning the perforations or liner slots is expected to improve flow and increase the production. However, historical data show that not all chemical treatment jobs yield the same level of success. The problem addressed in this paper is using a data-driven predictive modeling approach to make recommendations for candidate well selection using chemical treatment based on past performance of chemically treated wells. This paper presents the development of a fully customized machine learning tool that has been designed as a decision support tool for a particular task for non-data scientist reservoir and production engineers. The architecture of this tool includes independent modules that can be re-trained in the future once additional production data becomes available from the fields. We attribute the success of this research to the fact that there was active collaboration among domain experts. Critical physical properties were included in this study, including chemistry studies of chemicals and limited field trials in validating the developed model. The resulting machine learning model can be used by field engineers for planning chemical well treatments given the continuously updated past performance data.