Information-theoretic sensor planning for large-scale production surveillance via deep reinforcement learning

Abstract

Production surveillance is the task of monitoring oil and gas production from every well in a hydrocarbon field. Accurate surveillance is a basic necessity for several reasons that include improved resource management, better equipment health monitoring, reduced operational cost, and ultimately optimal hydrocarbon production. A key challenge in this task, especially for large fields with many wells, is the measurement of multiphase fluid flow using a limited number of noisy sensors of varying characteristics. Current surveillance practices are based on fixed utilization schedules of such flow sensors, which rarely change over time. Such a passive mode of sensing is completely agnostic to surveillance performance and thus often fails to achieve a desired accuracy. Here we propose an active surveillance approach, underpinned by the concept of value of information-based sensing. Borrowing some well-known concepts from Markov decision processes, reinforcement learning and artificial neural networks, we demonstrate that a practical active surveillance strategy can be devised, which can not only improve surveillance performance significantly, but also reduce usage of flow sensors.