Reactor runaway aware RL Agents for safe reactor operation

Fricz, Balázs and Szatmári, Kinga and Németh, Sándor and Nagy, Lajos and Bai, Wenshuai and Kummer, Alex (2026) Reactor runaway aware RL Agents for safe reactor operation. Computers & Chemical Engineering, 212. p. 109704. ISSN 00981354

Preview

Text reactorrunaway.pdf Available under License Creative Commons Attribution Non-commercial. Download (5MB) | Preview

Abstract

The control of semi-batch reactors is a challenging task due to the fact that these systems are non-stationary and often non-linear. Thermal runaway may also occur without strict operator supervision and rigorous control systems, which can lead to critical safety incidents. Reinforcement learning (RL) can offer a solution to the mentioned issues for autonomous process control by handling non-linearities and learning the uncertainties of the process. We applied Twin Delayed Deep Deterministic Policy Gradient (TD3) agents with different penalty weight factors in the reward function and compared their control performance on a reactor case study. Assessment of a thermal runaway criterion was included in the training of the agents to make them aware of potential future operating conditions and to ensure safe and stable process behavior. The proposed approach shows that applying RL for non-continuous systems is promising due to the observed flexibility and ease of setting up the system.

Actions (login required)

View Item