.png){kind=icon}
.png){kind=icon}
Return
A Risk-Aware LNG Terminal Scheduling Digital Twin based on Deep Reinforcement Learning
.png){kind=icon}
PRE
AI
Y
J
C
J
DOI：10.1016/j.compchemeng.2025.109520.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
At LNG receiving terminals, the daily send-out target fluctuates with seawater temperature, the accumulated runtime of parallel units, and planned start-up or shutdown sequences. Operators must still decide which pumps or vaporizers to activate under time pressure and incomplete information. In practice, these choices often rely on experience and short-term intuition rather than systematic evaluation, which can lead to uneven runtime distribution, maintenance bottlenecks, and unnecessary energy consumption. This study asks whether such human decision gaps can be reduced within the actual physical and organizational constraints of a working terminal. We develop RALT-DT, a risk-aware learning and control digital twin that integrates deep reinforcement learning with process-level physical models. The “risk-aware” feature is embodied in three aspects: (1) all policy decisions are constrained by explicit mass and heat-balance equations and safety interlocks, ensuring that the control actions remain within certified operating envelopes; (2) the learning reward explicitly penalizes excessive switching, uneven runtime dispersion, and deviations from preventive-maintenance requirements, treating long-term mechanical wear and operational stability as quantifiable risks; and (3) the system continuously monitors plant–model mismatch and adapts its confidence weighting, so that recommendations are moderated when uncertainty grows. To make the solution practical, the plant’s many operating devices are grouped into four core classes—low-pressure (LP) pumps, high-pressure (HP) pumps, open-rack vaporizers (ORV), and submerged-combustion vaporizers (SCV). A two-time-scale roster generator translates continuous policy outputs into binary start–stop schedules that meet maintenance lock-out and switch-inertia constraints. The resulting framework forms a closed learning loop that is both deterministic and interpretable. A one-month on-site shadow test was carried out, in which the algorithm’s decisions were compared with real operator schedules under live conditions. The digital twin achieved an average electrical energy reduction of 9.4%, equivalent to a saving of about 754 MWh, without violating throughput or switching limits. When calibrated against plant telemetry and vendor performance curves, the model maintained consistent accuracy across varying load and temperature conditions. These results indicate that a physics-grounded and risk-aware learning framework can systematically enhance human decision quality in large-scale terminal operations. It converts intuition-driven scheduling into a reproducible and auditable policy that improves both electrical energy efficiency and asset reliability.
Journal
C
IF:
3.9
Papers:
8.1K
Citations:
1.7W
Organization
Cited Papers