Agile Development of Machine Learning (ML) for Conventional Artificial Lift Systems in the Middle-East

Abstract

Abstract The majority of Oman's southern fields are produced by beam-pumps which consist approximately 2,000 wells, globally beam pumps remain an extremely popular choice for secondary lift. Identification and diagnosis of beam pumps using dynocards is an expensive human visual interpretation process. It does not only require significant labor time but also requires deep expertise in the production technology domain. The development team had three goals: 1) use open-source analytics, 2) develop a Machine Learning application (ML) to solve business challenges and finally 3) foster solutions with significant value investment ratio (VIR). In this case, a proof-of-concept application was developed to automatically screen beam pump dynocards and identify abnormalities (e.g., electrical related failures) which cause improper operation of the well, leading to deferment, undetected by conventional monitoring systems, and/or mechanical damage. To address the above challenges, an analytics minimum viable product (MVP) was developed for pattern recognition which significantly assisted in automating (analysis of a 100 hundred wells with real-time data in less than 1 second) the visual interpretation process, increasing efficiency, and reducing maintenance activities due to missed early diagnosis. It detects current and future abnormal conditions that cause improper operation of the artificial system to deferment and potentially to mechanical damage. This new app identifies and highlights these wells so that operations & maintenance staff can focus their attention where it is really needed, improving their workflows and decision making. This paper outlines how applying Machine Learning (ML) along with the Scaled Agility methodology enabled the operator to develop an MVP and diagnose abnormalities on daily basis not raised by any other system. Of the 100 wells in the selected field around ~10% were identified with clear failures. This translated to approximate ~5 % improvement in lead indicator (prior to issues) detection projecting ~2.5 million USD in efficiencies and deferment reduction. The cost of development in CAPEX was 0 USD as the team developed this purely on Open-Source platforms that were license free and on their own without the need of third-party application or resources.