.png){kind=icon}
Return
A Practical MPC-Pretrained Reinforcement Learning Framework for Complex Process Systems
.png){kind=icon}
PRE
AI
DOI：10.1021/acs.iecr.6c00277.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reinforcement learning (RL) is increasingly studied for process control, but online exploration can violate safety and operating constraints. Prior work from our research group reduces exploration by pretraining a twin delayed deep deterministic policy gradient (TD3) agent via behavioral cloning of an offset-free linear model predictive control (OF-MPC) policy. This approach significantly reduces unsafe online interaction and can exceed OF-MPC performance. However, our large-scale simulations show that MPC imitation alone, when paired with a one-step quadratic reward, can still lead to non-negligible steady-state error. To address this limitation, we modify the online phase in two ways. First, we augment the MPC-style reward with an offset-aware shaping term that increases the learning signal near the set point. Second, we use a mixed-replay strategy that combines prioritized replay, a recency bias, and uniform sampling to balance adaptation and coverage under drift. The method is demonstrated on two systems: a multi-input multioutput (MIMO) styrene polymerization reactor and a large-scale ethylene-splitter distillation column simulated in Aspen Dynamics. In both nominal and disturbance scenarios, the proposed RL controller improves transient behavior and reduces steady-state error relative to the previous work.
Keywords:
Animal feed
Computer simulations
Hydrocarbons
Process control
Journal
I
IF:
0
Papers:
733
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSaveModel predictive control of nonlinear processes using neural ordinary differential equation models
PREAI
ShareSaveShareSaveShareSave
View more