TY - GEN

T1 - Deep reinforcement learning algorithm for wellbore cleaning across drilling operation

AU - Keshavarz, Sahar

AU - Elmgerbi, Asad

AU - Thonhauser, Gerhard

N1 - Publisher Copyright: © 4th EAGE Digitalization Conference and Exhibition 2024.

PY - 2024/3/25

Y1 - 2024/3/25

N2 - We propose a novel framework for real-time drilling operation planning updates using deep reinforcement learning algorithm, enabling drilling process reactions to be automated. The framework includes a decision tree algorithm to represent the environment dynamic changes based on the imposed actions parallel to a Gaussian process algorithm to quantify the safe operating window in real-time. Combining these two algorithms leads to mounting a Markov Decision Process (MDP) environment for a decision-making system.We demonstrate the effectiveness of our framework by implementing an off-policy deep reinforcement learning algorithm, using a deep Q-learning network to create experiences, and employing synchronous updates on the agent. Given the essence of reinforcement learning, the framework can be efficiently implemented for on-the-spot decision-making, allowing the driller to receive an effective sequence of actions considering company policies.Our algorithm achieves state-of-the-art performance on weight-to-slip hole conditioning operation, a wellbore cleaning operation after drilling a stand before connecting to the next pipe. The performance evaluation exhibits its efficiency in real-time operation overhaul, eliminating non-value-added activities. Our framework thus opens the door for automating the process based on the operating parameters obtained in real-time

AB - We propose a novel framework for real-time drilling operation planning updates using deep reinforcement learning algorithm, enabling drilling process reactions to be automated. The framework includes a decision tree algorithm to represent the environment dynamic changes based on the imposed actions parallel to a Gaussian process algorithm to quantify the safe operating window in real-time. Combining these two algorithms leads to mounting a Markov Decision Process (MDP) environment for a decision-making system.We demonstrate the effectiveness of our framework by implementing an off-policy deep reinforcement learning algorithm, using a deep Q-learning network to create experiences, and employing synchronous updates on the agent. Given the essence of reinforcement learning, the framework can be efficiently implemented for on-the-spot decision-making, allowing the driller to receive an effective sequence of actions considering company policies.Our algorithm achieves state-of-the-art performance on weight-to-slip hole conditioning operation, a wellbore cleaning operation after drilling a stand before connecting to the next pipe. The performance evaluation exhibits its efficiency in real-time operation overhaul, eliminating non-value-added activities. Our framework thus opens the door for automating the process based on the operating parameters obtained in real-time

UR - https://www.earthdoc.org/content/papers/10.3997/2214-4609.202439018

UR - http://www.scopus.com/inward/record.url?scp=85217620324&partnerID=8YFLogxK

U2 - 10.3997/2214-4609.202439018

DO - 10.3997/2214-4609.202439018

M3 - Conference contribution

VL - 2024

T3 - 4th EAGE Digitalization Conference and Exhibition

BT - 4th EAGE Digitalization Conference and Exhibition

ER -